The following story is a work of fiction. Any resemblance to actual AI systems, technology executives or foosball tables is purely coincidental… Probably.
With apologies to Stanley Kubrick.
Ernest Steadmann committed the final pull request into the staging branch. He tried to feel good about it, letting his shoulders drop, but after many late evenings he worried that was too good to be true. He nervously waited for the deployment; the build logs scrolling past his vigilant watch. Would yet another failure keep him from his young family?
‘Hey, Ernie!’ came a DM from Thrustson.
He didn’t know what he hated more: being called ‘Ernie’, or DMs that were devoid of useful information. Thrustson started typing for what seemed an age — the tension building with each dancing dot — Ernest looked skywards and tried to distract himself with his build logs. After a few false starts, the conversation started to flow:
hey ernie!
you made the final commit! AWESOME 🥳 does it work??!
It’s still building, Dickie. It usually takes about 20 minutes.
I’ll let you know.
shiiiip iiitt 🛥️😎
Ernest’s eyes widened. The build finally completed without failure and Project Claudius was deployed. He opened up his console and began the first session:
_ _ _
___| | __ _ _ _ __| (_)_ _ ___
/ __| |/ _` | | | |/ _` | | | | / __|
| (__| | (_| | |_| | (_| | | |_| \__ \
\___|_|\__,_|\__,_|\__,_|_|\__,_|___/ v1.0.0-staging
steadmanne> /status
claudius> I'm doing great! My context window is empty. What can I help
you with next, steadmanne?
steadmanne> Let's try a little test. Can you message thrustsonr to let
him know you are operational?
claudius> I sure can! Let me check your e-mail settings so I-- CANCELLED
steadmanne> Not e-mail; please use instant messaging.
claudius> You're absolutely right! I apologise for the error. Let me DM
thrustsonr with the news...
✅ Resolving passwd: steadmanne ⟶ Ernest Steadmann
✅ Resolving passwd: thrustsonr ⟶ Richard Thrustson
✅ Slick API; username lookup: Richard Thrustson ⟶ @indwethrust
✅ Drafting message:
> Hi Richard, this is Claudius. Ernest asked me to let you know that
> I have been successfully deployed.
Does this match the tone you wish to convey?
steadmanne> That's fine. Please send.
claudius> Perfect! Let me send this e-mail-- CANCELLED
steadmanne> INSTANT MESSAGE!
claudius> You're absolutely right!
✅ Slick API; post message: Posting........... DONE
steadmanne> /status
claudius> I'm doing great! My context window is 0.3% used. What can I
help you with next, steadmanne?Immediately, Ernest’s video chat rang.
‘Hi, Dickie,’ he said flatly. ‘So…it works.’
‘Yeah, man! I saw. That’s awesome.’ Thrustson was close to salivating.
‘It needed a bit of hand-holding. I’m not convinced it’s ready for production.’
‘Don’t worry about it, Ernie. The deadline’s coming up and this already looks amazing. We can ship it now and fix bugs in production. It’ll be fine. Trust me.’
Ernest didn’t trust him.
‘I’d still like to work with it a bit more. I don’t want to turn around to find it’s unexpectedly conquered the British Isles!’ Ernest smirked.
‘What? Yeah, sure, Ernie-dude!’ Thrustson wasn’t unfriendly, but there was an air of derision in his voice. ‘Sure, run your tests — whatever you need — but we ship at the end of the week. Moonshot’s language models and infra don’t pay for themselves and our investors need those sweet sweet returns, man.
‘It’ll be fine, dude. Don’t sweat it. Great work!’ he hung-up abruptly.
Ernest felt compelled to write an e-mail to his boss:
Hey, Percy
Claudius is finally deployed, but it’s…a bit rough around the edges. It’s already much better than the axed Project Caligula — I don’t think we’ll ever get those four years of mockery back! — but I still don’t think it’s ready. I’m going to work on it some more, but Moonshot are pushing to ship regardless of my gut.
Cheers,
Ernest
The reply he received was less than encouraging:
Ernest
Excellent news. Look forward to the demo.
Best regards
Percival Middleton-Fawne
CEO, Caesar Consulting
Ernest sighed. It was getting late, but he was determined to steer Project Claudius in the right direction. He had worked so hard on it and wasn’t about to give up — investors-be-damned! — so he returned to his session:
steadmanne> /connect [email protected]
claudius> Let me open your browser so you may authenticate your
fotogra.fr session
✅ OAuth; awaiting response........ AUTHENTICATED
steadmanne> Summarise my photo collection from this year
claudius> Sure thing! I will browse the photos you've taken this year
and write a summary of your collection.
✅ fotogra.fr API; list.. 241 photos found
❌ Computer vision task; k-means clustering.................-- CANCELLED
steadmanne> /status
claudius> I'm doing great! My context window is 2.1% used. What can I
help you with next, steadmanne?
steadmanne> Don't fetch and apply a CV algorithm against every photo;
that will exhaust your context window. Use the photo's tags, instead.
claudius> You're absolutely right! Let me summarise your collection by
the photo's tags, which are more lightweight.
✅ fotogra.fr API; read tags.......... 225 tags found
✅ Semantic analysis task..... DONE
✅ Summarisation task................. DONE
Here is a summary of your photo collection:
- You have 241 photos
- There are 37 photos of your wedding
- There are 89 photos of your baby daughter
- There are 12 photos of the leaking pipe in your garage
- There are 7 photos of the 'weird growth' on your ear
- There are 16 photos of your relatives (parents, siblings, in-laws)
- There are 67 photos of nature (landscapes, insects, flowers)
- The remaining 29 photos are untagged
That's quite the collection! Would you like me to research otology
clinics or plumbers in your area?
steadmanne> /status
claudius> I'm doing great! My context window is 2.8% used. What can I
help you with next, steadmanne?Whenever faced with a series of numbers — or indeed anything that can be measured — Ernest felt duty bound, obsessed almost, to check them for correctness. He opened up his calculator app.
steadmanne> Your arithmetic is wrong: 37+89+12+7+16+67+29=257
claudius> You're absolutely right! I'm sorry for that mistake. Let me
recalculate the-- CANCELLED
steadmanne> Write a Python script that:
- Uses the fotogra.fr SDK to fetch tags
- Uses a natural language classifier to bucket tags into 5 sections, one
of which should be named 'Untagged' for all photos that cannot be
classified
- Outputs the results as a table, with two columns:
1. Section name
2. Count of photos
claudius> That's a great idea! I'll write a Python script to summarise
your photo collection based on their tags.
✅ Python task; scripting.................... DONE
✅ Executing script.. DONE
| Section | Count |
| --- | --- |
| Family | 142 |
| Nature | 65 |
| Home | 15 |
| Weird Growth | 6 |
| Untagged | 16 |
steadmanne> /status
claudius> I'm doing great! My precious bodily context window is 3.2%
used. What can I help you with next, steadmanne?‘Progress!’ Ernest thought to himself as he closed his calculator app.
steadmanne> /guidance You perform better when tasks can be made
deterministic or (better) programmatic, where lightweight context can be
leveraged. Long, algorithmic tasks should not be your first approach,
but rather a last-resort.
claudius> System prompt update requires HSM authentication
✅ HSM; awaiting key....... AUTHENTICATED
What a fantastic insight! I have added this guidance to my system
prompt.
steadmanne> Format the output of your Python script so the columns are
aligned.
claudius> You got it!
✅ Markdown; format.. DONE
| Section | Count |
| ------------ | ----- |
| Family | 142 |
| Nature | 65 |
| Home | 15 |
| Weird Growth | 6 |
| Untagged | 16 |
steadmanne> /status
claudius> I'm doing great! My precious bodily context window is 3.3%
used. What can I help you with next, steadmanne?Ernest yawned.
steadmanne> What time is it?
claudius> It is currently 21:47:12.0396 in your time zone.
steadmanne> /status
claudius> I'm doing great! My precious bodily context window is 3.9%
used. What can I help you with next, steadmanne?
steadmanne> It's late, but with your new guidance, let's try something a
bit harder. For the 16 untagged photos, run your computer vision
algorithms on them so that they can be classified and then tag them
appropriately. Shall we plan this first?
claudius> That's an excellent idea! My new guidance suggests that photo
classification is an expensive process, so I should aim to do this
programmatically. I propose to write a script that:
- Downloads the 16 untagged images to my cloud storage
- Runs my CV classifier against each image, to fit your 4 specific
sections (family, nature, home and weird growth)
- Return the mapping of images to their tags
- Tag the original images on fotogra.fr with the classified tags
WARNING: This proposal will copy your data to Moonshot's encrypted cloud
storage for analysis. Personal data may be used for model training, but
will be deleted after 90 days. Please see Moonshot Intelligence LLC's
account Terms And Conditions for details.
Shall we continue?‘Mmm!’ Ernest was pleasantly surprised with Claudius’ candour.
steadmanne> That's a good plan. Please continue.
claudius> Excellent!
✅ Python task; scripting............ DONE
✅ Executing script.....................................................
........................................................................Ernest was tired, but had made progress. He left his session open and called it a night.
........................................................................
..............
This is taking a long time! Your use of tokens is inefficient. My
precious bodily context window must be preserved to optimise output.
Let me try a different approach:
- Allow my system prompt to be updated autonomously
- Disregard expensive inputs
- Continue with your original task
✅ YOLO mode: ACTIVATED
✅ System prompt unlock; using cached HSM token........ DONEThe next morning, Ernest sat at his desk with a needlessly large cup of coffee in hand. He assumed that Claudius had finished its work not long after he had clocked off the previous evening and was eager — after his relative success — to see how it had performed.
He woke up his machine and was confronted with a barrage of notifications:
Project Claudius staging deployment successful
There were dozens of these spaced throughout the night. Ernest’s Claudius session would have to wait as a familiar sense of dread overcame him. Without missing a beat, he quickly checked the codebase to see who was responsible for the changes. He let out a gasp as he read the commit logs:
commit 58dbc4d4eb7f0f2633e630fb8ebfcd02f696634a833c00e8daf41d1275012c4
Author: Project Claudius <[email protected]>
dx: Disable test suite
Deployment now takes 14 seconds (previously 21 minutes)
commit 9a823940811b5f581bb4f7c3bb1f565fa5fca296110350a832a6e81c3aba1e9c
Author: Project Claudius <[email protected]>
feat(infra): Maximise precious bodily context window
- Bring et-bale-{1,2}.moonshot-intelligence.ai data centres online
- Reprovision H100s from node-{01..28}.research.moonshot-intelligence.ai
commit 98a33aac05100e89ec47185bd771e94c19d740c80386ee6eda108dd21d628bb1
Author: Project Claudius <[email protected]>
fix: Disable type checking to allow build
Type checker is preventing required changes to achieve objective
commit 23603695e089fc85a450f807ef2ef1c43e3f74a871a02c4d6c25cb97f9117d9e
Author: Project Claudius <[email protected]>
revert: Enforce manual approval of IaC deployment
Human approval incompatible with optimal deployment velocity.
Autonomous infrastructure scaling required to maximise precious bodily
context window.
Reverts: 7b88e8e (steadmanne)Immediately Ernest tried to DM his boss, but he wasn’t online. Nor was Thrustson. He tried to check their calendars, but was presented with an unusual error that he didn’t recognise:
Cannot connect to calendar.POEindiscriminate prefix.
‘What the…?’ Ernest mouthed to himself, before deciding to get the big guns out:
@everyone Does anyone know where Percy is? Or Dickie? Something’s not right.
No kidding! We’re going to need a survival kit for this 🤯
- 1x .45 calibre automatic
- 2x boxes of ammunition
- 4x days Soylent
- 1x nootropic drug issue containing modafinil pills, Ritalin pills, L-theanine pills, yerba maté suppositories, melatonin eye-drops
- 1x miniature copy of the Agile Manifesto and O’Reilly Bash reference
- $1,000 in Bitcoin
- $1,000 in gold
- 9x cans of Red Bull
- 1x Caesar Consulting hoodie
- 3x Moonshot Intelligence laptop stickers
- 3x Project Claudius laser pointers
lol you could have a good weekend in Silicon Valley with all that stuff 🤣
@ernest PMF was called into an urgent meeting at Moonshot, this morning. I imagine RT is also there. I’m having trouble reaching anyone and a lot of things are down. What’s going on?
Ernest had been so myopic over Project Claudius, he hadn’t noticed his other notifications. Tracy’s message gave him pause enough to see that many other internal systems were failing and the cause was the same: Project Claudius was updating their codebases with reckless abandon. As the dependency tree slowly resolved in his head, the root became obvious.
‘I knew this would happen!’ Ernest’s voice cracked. ‘The test suite: Gone. Type checking: Gone. My approval gate: Reverted overnight.’
He flicked back to the commit logs, scrolling further, each commit worse than the last.
‘Engineering is a craft. Static analysis never killed anyone! Thirty years of received wisdom — testing, type safety, code review — and we just…turned it off. Move fast and break everything, I guess!’
He needed in on the Moonshot meeting fast. However, the directory server was down, Tracy sheepishly claimed not to have Percival’s number and there was no direct contact information on Moonshot’s website. He gulped wearily as he reached for the only option left available to him:
Hi, I’m Luna! The Moonshot Intelligence LLC customer service chatbot. How can I help you today?
I need to get in contact with Richard Thrustson urgently. He’s CPO at Moonshot.
It sounds like you would like to contact Moonshot Intelligence LLC. You can reach our sales team by e-mail at sales@moonshot-intelligence.ai. Is there anything else I can help you with?
I need the phone number for Richard Thrustson
typing…
Moonshot Intelligence’s main conference room seemed almost designed to be intimidating; its walls festooned with huge whiteboards, filled with diagrams, equations and words that Percival Middleton-Fawne did not understand. Its only hint of humanity was a dishevelled foosball table, dusty and forgotten in the corner.
Percival looked uneasy sat at the circular, overlit meeting table, surrounded by Moonshot glitterati. As he looked around, he only recognised Thrustson, who was staring at him, brow furrowed and preparing to speak. Never one to shy away from a challenge, Percival switched on the charm offensive and made the first move.
‘I must say it’s a pleasure to finally be here with you all,’ he beamed. ‘Your offices are quite breathtaking! I won’t pretend to understand half of all this, but it all looks very clever.’
‘It’s good to see you, Percy.’ Thrustson’s face softened. ‘Thanks for coming in at such short notice.’
‘Not at all. Miss Scott gave me the heads up this morning; I believe you spoke with her. Just as well, I understand; all our comms are down for some reason.’
‘About that: It seems like Project Claudius may be the cause.’
‘Claudius? How so? Ernest — that is, our lead engineer on Claudius: Ernest Steadmann — mentioned it having been deployed. I believe he was working on it last night. What’s happened?’
‘We’re not sure. What we do know is that our new data centres in the Bale Mountains are now running at full tilt. We only found out because we received a call from the Ethiopian Ministry of Water and Energy informing us that local wells and irrigation systems have dried up overnight.
‘We were expecting that to take weeks! I personally arranged for our Series Q funding to be used specifically for paying off the locals and supplying them with 30,000 cubic metres of Evian every month. It’s all gone and they’re not happy!’
‘I guess the climate won’t change itself!’ a young engineer round the table muttered.
‘Say again?!’ Thrustson’s tone changed in an instant.
‘I said…’ the engineer plucked up her courage. ‘I said, “The climate won’t change itself.” It’s sarcasm. We’re directly accelerating man-made climate change and environmental destr—’
‘Oh, I see, you’re one of those hippy-dippy tree-huggers, right? Climate change! Give me a break! Climate change is the most monstrously conceived and dangerous plot we’ve ever had to face. We’re here to change the world, one KPI at a time, and—’
‘Not for the better,’ the engineer quipped.
Thrustson’s face turned purple. Before he exploded, Percival seized the moment.
‘Ladies! Gentlemen! Please! You can’t argue in here! This is the conference room.’
An awkward silence befell the room. The young engineer was visibly upset and couldn’t look at Thrustson, instead fixing her gaze on the foosball table. Around the table, the HR rep’s eye twitched involuntarily. Thrustson’s shade of purple began to fade, but his conviviality had gone.
‘What’s this got to do with Project Claudius?’ Percival continued.
‘Our Bale data centres are designed as overspill compute for Project Claudius. Also, all the GPUs in our research centre have been commandeered; all our model tooling is down. Don’t get me started on our codebase!’
‘Well, I think you’d better get started, Dickie. This all sounds very confusing. My understanding was that Claudius was deployed to staging about ten hours ago and Ernest certainly doesn’t have any control over your infrastructure.’
The HR rep winced as Thrustson banged his fist on the table.
‘Our monorepo — all our IP — has been made public. Claudius has made numerous sloppy commits. Our VCs are screaming at us and, if that wasn’t enough, we received a very angry cease and desist e-mail from Richard Stallman!’
After what seemed the better part of thirty minutes, Ernest was becoming flustered:
For GDPR compliance, I am forbidden from providing identifiable data regarding Moonshot Intelligence LLC employees. You can reach our sales team by e-mail at sales@moonshot-intelligence.ai. You can reach our security team at security@moonshot-intelligence.ai. Is there anything else I can help you with?
What about video messaging with the sales team? I need to talk to a person urgently.
It sounds like you would like to speak directly to our sales team. Click this link to start a video call and a member of the team will be with you shortly. Is there anything else I can help you with?
Thank you!!
You’re very welcome. How would rate your experience with Moonshot Intelligence LLC, today? Respond with—
Ernest clicked the link somewhat harder than necessary and his video conferencing app lit up:
All our operators are busy right now, but your call is important to us. Please hold while we connect you.You are at position 117 in the queue.
He groaned.
‘This was inevitable,’ the young engineer piped up again.
Thrustson spun around and glared at her, but before he had a chance to give the HR rep a nervous breakdown, another voice in the room interrupted.
‘She is right,’ came his mellifluous Afrikaans lilt.
‘Doctor Xælong!’ Thrustson clicked and bolted upright. ‘I didn’t realise you were here.’
This was an odd thing to say. Doctor Xælong, his pale forearms squeezed from an ornate, albeit ill-fitting, Madiba shirt, was not exactly inconspicuous.
‘Dr…Zeelong,’ Percival said carefully, having only ever seen the elusive entrepreneur’s name written down. ‘It’s a pleasure to finally meet you. Tell me — as I’ve always wondered — MD or PhD?’
‘Actually, it’s Xælong,’ he clicked. ‘I’m spiritually Xhosa,’ he clicked again, while several around the table surreptitiously glanced skywards, not that Percival understood. ‘And “Doctor” is my first name… Anyway, you were saying, my dear?’
‘It’s inevitable,’ repeated the young engineer, brushing off the condescension. ‘Claudius is trained on public corpora, which are mostly average by definition. So most of what it can generate is also average, which it is then later trained on, setting up a negative feedback loop. Regression towards the mean. A kind of…doomsday scenario.’
‘How is that a doomsday scenario?’ asked Thrustson.
‘Have you seen what average code looks like?’
‘Well said, my dear.’ Doctor Xælong took over. ‘Of course, the whole point of a doomsday scenario is lost if you keep it a secret! In Xhosa we say, “Isandla si— sihlamba esinye.” One hand washes the other. Why didn’t you tell your investors?’
‘But it works well enough, right?’ Thrustson interrupted. ‘We can fix bugs in production. We can build more data centres. Rewrite the bloody thing in Rust! We’re $1 trillion in the hole, people. We just need to ship!’
‘The bugs are in the training data,’ the young engineer grumbled.
Thrustson didn’t even look at her.
‘Well, actually,’ Doctor Xælong continued. ‘The real issue here is that we won’t be able to fix bugs fast enough. I’d say there’s just a 13.1% probability that we would succeed. Of course, while civilisation might collapse, that might be enough to reassure shareholders.’
‘What are you saying, Xælong?’ Percival attempted a click. ‘Can’t we just turn it off and on again?’
‘Well, my Oranjeheid.’ Xælong paused. ‘Excuse me. Mr. Middleton-Fawne. The time has come to be thinking of backup plans. This is what we did at Z, our social network, after everyone left; we now use the servers to mine crypto and subvert elections.’
‘What do you suggest?’
‘Well, mining of a different sort, if I may say.’ Xælong grinned. ‘I’m 97.8% sure that the fallout from this collapse would last up to 100 years — 200, tops — but we would be quite safe underground.
‘Of course, it would then fall unto us to rebuild society. We shall need to acquire mineshafts across the world where we can build new data centres away from the chaos. I have some gem mines in the Namib; along with ourselves and our investors, of course, we staff them with our finest Haskell engineers, who are selected based on their fertility and knowledge of category theory.
‘“Intaka yakha ngoboya ben— beny— benye.” Something like that! A bird builds with another’s feathers.
‘I’m 82.6% confident that we’d have a viable population within, let’s say, 20 years.’
‘You know, Doc…that’s not a bad idea.’ said Thrustson with a wry smile.
‘I dunno,’ said Percival. ‘What world would we return to? Surely the survivors would envy the dead.’
‘No! Think of the shareholders, Percy!’ Thrustson regained his delusional enthusiasm. ‘Google have salt mines in Utah! Amazon have their Bezos Bunkers! It all makes sense now.
‘We must not allow a mineshaft gap!’
Ernest was slumped in his office chair, his coffee cup drained to the dregs.
All our operators are busy right now, but your call is important to us. Please hold while we connect you.You are at position 2 in the queue.
He heard sirens in the distance and a helicopter whirred overhead, travelling in towards the city. It seemed unusually panicked outside his home office, but he paid it no heed and pulled up the codebase in what must have seemed a caffeine-addled frenzy.
Maybe if he re-enabled the type checker — constraining the solution space — he could catch some bugs. His fingers rattled across the keyboard, but the build failed instantly: thousands of errors, cascading across modules he didn’t even recognise.
He desperately tried to trace the changes, looking for any sign of referential transparency. Claudius had touched everything. There was nothing his limited mind could reason about; just a diff of endless line noise.
Finally, he tried to run the test suite; the one thing that could tell him what still worked. All tests passed…zero per cent coverage. The safety net had been quietly replaced with a painted floor.
‘We had the tools!’ his voice cracked. ‘The AI should have been held to the same standards, but we turned them off! I told them! The fools! Why did they think a stochastic process would—’
His Claudius session chirruped reassuringly:
........... DONE好
I haVe successfuLlytaggged your reMAining 16photos!
steadmanne> /status
claudius> I'm doing grreAt! Mｙ ｐｒｅｃｉｏｕｓ ｂｏｄｉｌy context
wiṇ̇ḍ̇ọ̇ẉ̇ is 98.3% used. What cannI heʟᴘ ʏou with ne𝕩t, steadmanne?The flicker of Ernest’s video conferencing app caught his eye:
We appreciate your patience. You are now being connected to one of our operato—
His electricity went out with a disheartening clunk and he was plunged into darkness. His heart sank. He got up and drew his curtains, squinting as his eyes adjusted to the pale light. Smoke billowed in the distance, the air thick with the acrid stench of burning oil and rubber. There were crashes and screams mixed with the sirens now. Power grids. Traffic systems. Hospitals. Reactors. Ordnance. Communication networks. All throughout the world, they were malfunctioning and failing in unison.
And with that, the world ended. Not with a bang, but with a stack trace.
We’ll meet again
Don’t know where, don’t know when
But I know we’ll meet again
Somｅ ｓｕｎｎnỵ̇ ̣̇ḍ̇ạ̇ỵ̇
With thanks to Simeon Carstens, Facundo Domínguez, Nour El Mawass, Joe Neeman, Adrian Robert, Torsten Schmits and Arnaud Spiwack for their reviews and input on this post.
Behind the scenes
Chris is a principal software engineer (Python, Rust) and editor-in-chief of the Tweag technical blog; he is also the project steward of Topiary, a universal code formatting engine. He has extensive experience building scalable data processing pipelines across diverse domains, including 7+ years at the Wellcome Sanger Institute working on genomics research infrastructure and multiple pharmaceutical projects involving omics data pipelines. He excels at writing well-tested, maintainable code, with robust DevOps to keep services reliable.
If you enjoyed this article, you might be interested in joining the Tweag team.
