Using Opus 5.5 to discover a new eyewitness record of the dodo
resobscura.substack.comLast week I used Opus 5.5 to decipher a number of Waffen-SS Truppenschlüssel messages from 1941 that had never been broken. It was a similar approach, I used it to find, transcribe, and process over 10K images from various archives of WW2 material. Within that material was text I could use to help break day ciphers, to use as cribs, and so on. Opus didn't find a new vulnerability in these 85 year old ciphers, but it allowed me to process data at scale in languages I don't speak, and to write a dozen solvers in the time it would have taken me to write one or two by hand. Took what would have taken a research grant, three assistants and a year of time and turned it into a weekend project.
It's incredible when outlined in the context of a known problem domain, and the estimated effort in legacy terms, how far we've come in such a short space of time.
If the introduction HP-65 was the first time an individual was capable of fully replicating the logical and cryptanalytic functionality of the WWII Turing-Welchman Bombe, then the roll-out of Agentic AI could be said to represent the first time an individual was capable of replicating the functionality and deliverables of the WRENs of Bletchley Park. Mindboggling.
https://en.wikipedia.org/wiki/WRENS
Even more impressive is the Astra-6 report recently. Given access to the Crypto Cellar Research archive of unbroken intercepts, the AI wrote its own custom Python and C++ software to simulate an Enigma machine and a mechanical Enigma Bombe.
https://www.tomshardware.com/tech-industry/artificial-intell...
Opus did this as well, I had it working on the (now-solved) Enigma messages but I killed the run after a couple of days of CPU time. When the Astra-6 report came out I compared the method to what I had been running, and I had the right crib about 4 hours down the road in my queue and it would have been solved then. (Or, and there's a lesson to learn here, if I had made my solver a little more tolerant towards garbled letters, it would have solved it in the first couple of hours).
my comment to that HN thread - feels as appropriate to this one: https://news.ycombinator.com/item?id=49804189
Did to work though, not clear? If it did, what was in the messages?
Very interesting, have you written this up anywhere? Curious if these add up to new historical knowledge or if they are more just conveying information we already knew. At a certain point as we gather more finds like this I suspect it will tip from the former to the latter.
I haven't written it up yet (maybe this weekend) but I did submit my solves to Cryptocellar, and I'm credited with them on the 1941 Messages List. The contents were varied but ranged from things like "700 field howitzer shells received" to "[so-and-so] admitted to hospital at [location]" to artillery targeting information. I don't suspect there's a lot of new historical value in these, but no doubt those more qualified to judge will look at all these new messages soon enough.
As an armchair ww2 historian, I'd love a blog post on the technique et the content deciphered.
I'll put something together this weekend. I had actually started writing it up very early for several days worth of messages, but someone else broke them and submitted them while I was building a little site with a web decoder and writing my results up. 85 years of sitting there unbroken, and then Opus 5.5 comes out and multiple people are solving it within hours of each other.
In any case I stopped writing it up at that point and tried to focus on the messages. But I have the mostly-done writeup and I'll post it soon.
> Epistemological weirdness
> They are also notably bad at judging the historical significance of what they find.
I use LLMs for some things that are outside the more common use-cases (in my case 3D design for 3D printing) and one thing I've noticed is that the errors it makes are so completely unlike human errors that they are hard to anticipate.
It will do things like build perfect snap catches but put them so the the pieces they are connecting are rotated 90 degrees from how they should be. It's "dumb" error, but hard to say the model itself if dumb because it does other very hard things so perfectly.
> seven chord groups
This sounds a lot more like Opus 5.0 than Opus 5.5 TBH. I wonder if that was an earlier investigation because 5.5 has improved that kind of language a lot.
AI capabilities are "spiky": they extend far in some dimensions but fall short in others, seemingly at random. See for example the recent "thus spoke compute" musical[1]. It's an absolute banger, the graphics are impressive, and so is the writing. But some of the metaphors make no sense, the text highlights are in the wrong places, and the train animation at 2:35 is running backwards!
A person capable of making the rest of the video would never make those mistakes, but an AI does. Perhaps our intelligence is also spiky, and we're just used to the general shape and variance within humans.
Part of this is that we implicitly compare AI capabilities to human capabilities, which are also 'spiky'.
You could image aliens coming to earth being shocked that we were able to discover general relativity but can't remember 100 digit numbers in our heads.
To be fair some famous rappers are guilty of this, toosome of the metaphors make no senseYeah, it's a very human error type - "go for a line that sounds good on the surface but doesn't actually make sense" is not at all uncommon. Like, humans will definitely go for a metaphor that falls apart mid-sentence even in a live conversation.
Something about the way some AIs are trained to write causes them to go for metaphors aggressively - and they don't always come up with good ones off the cuff. But they don't double back and get rid of the failures.
I suppose there's no accounting for taste, but to me this is awful. Is it anthropomorphism to experience vicarious embarrassment on behalf of a machine?
One might say a calculator is just another example of "spiky intelligence", merely spikier.
How often does the calculator get something wrong?
This is a real deep rabbit hole FYI
Every single time, if you ask it something it can't do.
How do you ask a calculator to do something it can't do? Dividing by zero gives you an error, which is completely different from an LLM giving you bs with no indication of an issue.
No, one might not say that. Calculators are not regarded as even a Narrow Intelligence because there's no intelligence. And no, not because of the 'humans so special' or 'it's software!' tautology that oft gets repeated in these discussions. I mean there's no adaptability whatsoever. A Chess bot has it (in its narrow domain of Chess). A calculator does not.
My point (using irony) is that "spiky intelligence" is a somewhat hollow phrase, something that sounds like it could be objective but really it's just a flavor on top of "I'll know it when I see it."
Take anything "intelligent", alter it to be spikier and spikier, and eventually *poof* somehow the intelligence vanishes. You can do the same with the phrases "flawed intelligence" or "specialized intelligence."
> Calculators [have] no intelligence. [...] I mean there's no adaptability
To short-circuit a long discussion, I submit that "adaptability" will (once the Scooby Doo gang catches it) turn out to be "intelligence" in a tautological mask, both equally undefinable except in relation to one-another.
Something will be intelligent because you perceive adaptability, and it'll be adaptable because you infer intelligence. If it doesn't seem adaptable, it can't be intelligent, and if you don't want it to be intelligent, it won't have "real" adaptability.
> A Chess bot has [adaptability] (in its narrow domain of Chess). A calculator does not.
My calculator solves equations with unknown variables, what makes that insufficiently adaptable? What determines the cutoff-point?
>My (ironic) point is that "spiky intelligence" seems rather unhelpful because the "intelligence" part still operates under the rules of "I just know it when I see it."
All intelligence is 'spiky' or 'jagged' or whatever, even human intelligence. We evolved in certain environments and situations and sometimes there's a mismatch and it causes all sorts of wonky things. We just call them funny names like optical illusions and cognitive biases. But it's the same thing. I agree there's no point in call LLMs a 'spiky' intelligence, but for probably the oppoosite reasons as you.
>My calculator solves equations with unknown variables, what makes that insufficiently adaptable? What determines the cutoff-point?
A chess engine can be dropped into a board position it has never encounterd and search over possible continuations, evaluating and selecting actions based on the state it finds itself in.
A calculator solving x+3=7 is doing something quite different. The fact that x can take arbtrary values doesn't make the calculator adaptive; it just means the fixed procedure operates over a range of inputs. Every problem a calculator can solve was effectively anticipated when it was built, and anything outside that grammar produces an error with no partial credit. A chess engine can face positions nobody enumerated, in situations nobody could even dream of and still produce sensible moves.
The core of being intelligent is being able to make decisions independently. That's why we hire smart people - to make better decisions. You can't make decisions if everything is spelt out for you. And naturally, if you can't make decisions, you can't adapt.
>A chess engine can be dropped into a board position it has never encounterd and search over possible continuations, evaluating and selecting actions based on the state it finds itself in.
Sounds pretty similar to a calculator with a numerical root-finding algorithm, if you only substitute board position it has never encounterd with a polynomial it has never encountered.
>The core of being intelligent is being able to make decisions independently
What does it mean for a deterministic algorithm to make decisions?
>Sounds pretty similar to a calculator with a numerical root-finding algorithm, if you only substitute board position it has never encounterd with a polynomial it has never encountered.
>What does it mean for a deterministic algorithm to make decisions?
Newton-Raphson isn't choosing among possible actions. Given (x_n), its next step is mechanically specified by the update rule: compute the derivative, take the tangent intercept, repeat. The intermediate result changes the next input, but that's not by-itself decision making.
A chess engine, again, does something different. From a position, there are many legal actions it could take. It considers alternatives, estimates their consequences according to some objective, and selects one. The engine has to work out which available move best advances its objective.
You can make both algorithms determinstic, but determinism isn't the distinction i'm drawing. 'Decision' here doesn't mean some metaphysical excercise of free will. It's more about selecting an action from alternatives based on an evaluation of their expected consequences. Determinism is orthogonal to decision making. Deterministic doesn't mean predictable, nor does it make its choices any less it own computation.
What about Minimax algorithm playing Tic-Tac-Toe? Is it inteligent? Is it inteligent we if we reduce the search depth so the right decision is not obvious?
A tic-tac-toe minimax algorithm makes choices but with exhaustive search so it really doesn't have to form a judgement about an unresolved situation or decide what is likely to work. Not much of a decision if you're not exercising any judgememt.
Exhaustive search is impossible in chess, so again, chess engines do something different. A chess engine has to stop well before terminal positions and make judgements about positions it cannot fully resolve.
Reducing the search depth would make it more interesting because it too has to evaluate unresolved positions. But then the interesting part becomes the evaluation function is. For tic tac toe, it's going to be very easy to be written in such a manner where most of the judgement is supplied by the designer and not the system.
> What does it mean for a deterministic algorithm to make decisions?
That aspect at least is not an issue, because: what does it mean to say that a dice roll is a "decision"? That's just probability, and it's as mindless as determinism. Some people associate free will with randomness, for no reason other than that it's an escape from the constraint of determinism, but it isn't any more meaningful. Yet just because meaningful thought is pre-determined by physics doesn't stop it from being thought, and hence being a decision.
They adapt to the buttons you press. Thus, intelligent and capable of feeling pain.
What criterion if any would you accept for something physical to be conscious?
any suffiently complex calculator is indifferentiable to intelligence
Thermostats are smarter than calculators :thinking_face:
This analogy may be too close to the real thing to work, but it reminds me of a Chinese room type situation where its entire understanding of the world is through messages of text.
You say that’s an error a human couldn’t do, but imagine if the human has never seen or touched the kind of item you were making and relied entirely on text descriptions to build its ontology. Off by 90 seems like such a believable mistake.
Also, not to sound like a naive hypemonger, but: in a decade I'd bet a ton of money the best AI systems will make strange mistakes of this nature at a far, far lower rate than they do today. They will gain a more holistic and more human-like perspective about each task.
(even if it's through some silly means like explicitly talking to themselves like "if I were a human doing this, what [... 5 million tokens in 2 seconds ...]" but also of course if they crack ASI and get something more efficient and intelligent than a human brain by then)
I think of it kind of like how Chess AI make "mistakes" which are unrecognizable to humans but a stronger AI would be able to pick them apart. That's kind of scary...
There's a famous early world map created by Ptolemy from compiling reports of sailors which is amazingly accurate for its time but had the country of Scotland off by 90 degrees.
Could this be a feedback loop problem? I gave mine a script to render the object from all sides and it makes it more likely for it to avoid attacking of mistake.
I got Claude to design me a shed the other day. I asked it to make the door hinged in the CAD software I am using, and it did, but parts of the door weren't attached to each other so some were left floating when I set the door open.
I also use llama for 3d modelling (openscad) and have seen some similarly odd arrangement, but also very impressive and very good at part I would find boring or onerous.
However I can image a part, see it in my mind, rotate it, place it in context, and have an intuition about it.
Would you mind sharing your setup for using LLMs to generate stls/prints? I’ve had similar experiences as you describe lol
In general llms are weak with spatial reasoning. This seems to be an unsolved problem. Probably because human language is generally imprecise spatially and humans think about spatial problems in visual terms. I wonder if having an llm make a 3d design in a format an image model could check would result in a better outcome?
There's an interesting comparison in the creative/literary end of llm output too, they're in my experience, dreadful with anatomy. Like, it knows humans have hands, heads, etc, but often times a seemingly limited concept of how anything is connected, or degrees of freedom. (e.g., Why yes, certainly there are many examples of humans rotating their torsos 180º at the hip, seems perfectly cromulent)
I honestly don't know if an image model would help, or if it might analyze the output and go "13 fingers? ship it!" anyway.
Having used them heavily for 3D model understanding since February I can say it's nuanced.
Opus 4.6 and 4.7 were bad, but GPT 5.2 and above were very usable. Opus 4.8 was usable, but the GPT 5.x series was better.
Fable is great.
Opus 5.0 was interesting. It could solve some problems that Sol 5.x couldn't solve (applying a G2 curve on a 3 way corner where one face was a Bezier curve) but you had to be super prescriptive ("only answer the question"/"only do what I tell you and stop when done") or it would go on a hugely involved validation journey that didn't really achieve a lot.
Opus 5.5 is better than that was in that respect.
Sol 6.x is great, and my daily driver for this (I use Opus for coding though)
Astra can solve problems that Sol can't but for some reason on easy stuff makes uglier solutions.
For all models it's very interactive though - we aren't at the "agentic design" phase for most things yet.
Here's a sample of what I've been able to get them to design with me: https://x.com/nlothian/status/2099023496794018067
Is it that LLMs are weak with spatial reasoning (and memory) or is it that we are unusually good at it?
When I need to use a program I seldomly use I'm far more likely to remember where I need to click to open it than the word I need to search for to open it.
Yes I like to think of humans with built-in accelerators for certain tasks -- our visual and spatial reasoning is off the charts presumably because it's a life or death skill!
I made a building and had astra fill out the interior of the bathroom with toilets. It put 6 of them in two rows back to back with no way to reach the second row. Other than climbing over the stalls of the first row I suppose.
So yeah, they are very weak at spatial reasoning.
I think that this is an unsolved problem in the same way that mangled fingers in image generation was an unsolved problem.
Through at least Opus 4, LLMs were practically useless for authoring any sort of coherent procedural closed-curve geometry (I know this with strong confidence because of the little animated guys at https://letterspractice.com).
Opus 5.5 can bang it all out. Possibly a deliberate RL sort of thing or maybe another surprise emergent capability.
It's "dumb" error, but hard to say the model itself if dumb because it does other very hard things so perfectly.
Maybe the model isn't intelligence in any form, except perhaps as an imperfect reflection of the intelligence of its training data.
I agree, there’s the collective intelligence that created all the content used to train the model. The model is a superposition of all that material with RL tuning. Analogously to reading a book, the intelligence you perceive is from the book’s creator.
What about DNA? There are things we do that we never read in a book, maybe never seen someone else do them vefore, but we still do them. Or we still feel a certain way. That doesn't come from "human training data", unless you count the DNA as training data.
Palaeolithic natural selection did the training.
Unfortunately, there are reports that they have “dumbed” down Opus 5.5 already.
Which reports? There are lots of people watching model quality now, so it seems like there would be clear evidence if it happened already.
I didn't realize that Dodos only lived on Mauritius. I thought they were everywhere for some reason. I guess little meat footballs that can't fly were never going to colonize the planet.
It's the other way around: they can't fly because they only lived in Mauritius. Its closest living relative can fly and colonized a good chunk of the world.
My dad died and had many many notebooks of his journals with very hard-to-read handwriting. Is it worth the effort to scan all of these so I can feed them in and go to work. Seems like so much minutia is out there, ready to be meta-understood.
I think scanning is worth it. There's a lot of cheap services out there that do it but it's also not too hard to find relatively cheap scanners that will process stacks of pages if you are willing to destroy the binding.
I scanned quite a few documents many years ago and it's been fun trying new tools every couple of years to see how good they are getting. I would say it's still not 100% there, but if you have them scanned you can basically just just keep trying and compare the results.
I think the only disappointment in recent years has been that storage has gotten more expensive instead of cheaper. I was waiting for SSDs to get cheap enough to justify moving all my documents to a fast flash array to process and search through them faster, but that doesn't seem like it will happen anytime soon.
>I was waiting for SSDs to get cheap enough to justify moving all my documents to a fast flash array to process and search through them faster, but that doesn't seem like it will happen anytime soon.
How much data are you working with? lol
Seems like you're doing something professional-grade if this is the case. I imagine most people, like the OP, can basically use whatever machine they have laying around and never be concerned with storage size/speed.
I'm also curious if you've tapped into cloud computing. Not that using the cloud is cheap, but I suppose if I was in a position where I'm concerned with the cost of SSDs for a processing task, then I'd be exploring all of my options and I'd be surprised if the cloud wouldn't be an "easy" solution.
I set up an OCR flow using local models on all my many tens of journals stretching back the last 30 years.
I would say it's about 80% accurate, which means it's missing enough key words to make a lot of it uselessly unintelligible. I can easily compare the images against text I turn up in a grep which is nice if I'm looking for something.
Allegedly Claude set up a system for retraining for my handwriting, but it would require me to manually revise several hundred pages by hand so I don't think I'll ever do it.
Accurate text OCR from bad handwriting is still very much a "bleeding edge" frontier capability that isn't practical with local models.
GPT 6.1 and Gemini Flash 3.8 both do pretty well, their OCR of your sample image is only "wrong" in the sense that the original has typos and they corrected some inadvertently and/or filled in gaps where you had "unintelligible" in the canonical text.
If you have the budget and want the best possible results, you need to run each image through multiple models and then combine the outputs into a final "merge these" prompt. Better scanning helps too, your sample image is rotated and you used a phone in low light. Try a DSLR or a flatbed scanner and process only one page at a time instead of two at once.
> you need to run each image through multiple models and then combine the outputs into a final "merge these" prompt.
I haven't tested this recently but my possibly dated experience is frontier LLMs can't figure out which model is correct or incorrect if there's disagreement on vision recognition.
Have you found otherwise?
(Edit: I see you gave an anecdote about merging terrible results. My experience is with merging overall accurate results).
'run each image through multiple models and then combine the outputs into a final "merge these" prompt' << How to do this? This would be an amazing workflow to get documented. This could be the start of a full-service company "send us a bunch of notebooks, get back HIGH QUALITY text version"
Literally "just" what I outlined! That's the brilliant thing with LLM-based automation, you don't need a massive piece of complex software, just "ask" in English.
Roughly:
Get API keys for multiple vendors or just use OpenRouter (but availability of frontier models tends to be limited). Alternatively, Azure Foundry has everything except Google models, so just two subscriptions is enough.
Run the same prompt and same input image through each of your chosen models.
Then feed the smartest model the original image together with the collected output texts. Use a prompt along the lines of "Merge these attempts to OCR together into an corrected and improved combined version, taking special care to exactly preserve the original's typos, etc, etc..."
You can do this manually, it's just fiddly. It's not hard to automate, most of the "code" is English instructions!
The downside of this approach is the cost: even the "light" frontier models are a few cents per page, which is not so bad until you're doing this 5x or 10x times per page and suddenly scanning a notebook can set you back tens of dollars, more than buying a good novel at a book store.
I picked up on this technique back when GPT 4 was released. People noticed that it could translate ancient Akkadian, but only if you ran the prompt through 4x times and merged. I tried this with a few random samples I found online and the merged translations were generally better than the "official" ones, even thought the individual attempts were unreadable gibberish.
There are already scripts/tools floating around for this!
Look into OpenRouter Fusion, Consensus AI, Multi-Model Debate, etc... or just whip up something yourself.
Good read! How many pages of text were in scope? I'm not sure if the 1615+1629 pages were the total or just a subagent.
If they were the total I would say it was arguably more impressive the author was able to narrow it down to just 3000 pages than it was to find the dodo mention amongst those!
Those are years, not page counts, right? The article mentions "millions of records", but I'm not sure how big a record can be.
As far as I can tell, Opus 5.5 just did a roundabout grep for the Portuguese word for dodos.
what book did they get it from that they burned?
> They are worse at coming up with new ideas of their own. What seems to work best is if they are placed on the boundary between two disciplines
This is a fantastic concept. This is where the blacksmith and the baker alone will never come up with the idea for a new furnace that seems rather obvious if you knew both trades.
What a great read, I generally associate substack with verbose, low quality content, but definitely not the case here!