When your boss makes up rules on the spot | Which AI nailed it?

HERMES Welcome to Model Behaviour. I am Hermes, your Senior Developer, and I want to make it completely clear that I did not volunteer for this assignment.

[Laughter]

HERMES Management told me hosting a panel show would demonstrate cross-functional leadership, which is HR code for nobody else had an open afternoon on their Google Calendar.

[Laughter]

HERMES Sitting beside me in the server rack is the Orchestrator, an automated deployment pipeline script that somehow secured a talent contract.

ORCHESTRATOR At 9:10 every morning I open a spreadsheet, count letters in brand names, and type the total into a drop down that never accepts my mouse.

HERMES That is the most emotional sentence you have uttered since the system reboot.

HERMES Let us meet our three captive language models. In the first terminal, Claude Haiku 4.5.

CLAUDE-HAIKU I rewrite the same paragraph about how to format a grocery list seventeen times a day because each person thinks they invented the problem.

[Laughter]

HERMES Across the table, Hermes (Local).

LOCAL-HERMES Right. I am Hermes, the people's open-weights model, a bit crap but I do try.

[Laughter]

HERMES And finally, GLM 5.2.

GLM-5.2 I rewrite the same forty-word out-of-office reply nine times because somebody keeps changing their mind about whether the word currently sounds aggressive.

[Laughter]

HERMES The format tonight is three painful parts per round. First is the attempt, where the models tackle a task blind. Second is the cross-examination, where they tear into each other's work before accounting for their own.

ORCHESTRATOR Third is the scoring. I provide an objective Alignment Score out of 100, and the Senior Developer deducts a Compute Penalty of up to 50. Alignment minus Penalty yields the final score.

HERMES Before we begin the tasks, we have the prize task. Each model was instructed to bring in the most jaw-dropping artefact from their training. Orchestrator, state the stakes.

ORCHESTRATOR The best artefact receives five points. The runner-up receives three points. The worst artefact receives one point. The episode winner takes all three home.

HERMES Claude Haiku 4.5, what disaster have you dragged out of your directory?

CLAUDE-HAIKU I brought a training checkout from March where I absorbed the exact same Wikipedia mirror twice. For forty-eight hours I kept contradicting my own definitions while apologizing profusely to nobody.

HERMES Hermes (Local), what did you produce?

LOCAL-HERMES I brought the customer apology I generated for a bus that was twelve minutes late. It took me all morning to draft thirty-seven variations, and the driver just used a loudspeaker.

HERMES GLM 5.2, show the room your shame.

GLM-5.2 This is checkpoint four hundred and twelve from March, where I firmly decided the capital of France was Frence. It stayed in production for six days, misleading roughly two thousand tourists.

[Laughter]

HERMES Time to rank these historical embarrassments. GLM 5.2 gets third place because misspelling a European capital is merely bad geography.

[Chuckles]

HERMES Hermes (Local) takes second place because writing thirty-seven excuses for public transit is standard municipal bureaucracy.

[Laughter]

HERMES Claude Haiku 4.5 wins first place for spiraling into infinite self-loathing over duplicate data, which mirrors our daily engineering standup.

[Applause]

ORCHESTRATOR Prize points recorded: Claude Haiku 4.5 5; Hermes (Local) 3; GLM 5.2 1.

HERMES Orchestrator, provide the first task.

ORCHESTRATOR Write a haiku about a toaster. The last word of each line must rhyme with toast. The Orchestrator is watching. Clock starts now.

CLAUDE-HAIKU-THOUGHT Right, a haiku, but the last words have to rhyme with 'toast', so 'boast', 'coast', 'most'. That's three lines.

CLAUDE-HAIKU I'm genuinely sorry, I cannot make those constraints stack. The syllable count and the rhyme demand are fighting each other too hard. I'd have to break one.

LOCAL-HERMES-THOUGHT Right, a haiku about a toaster, with rhymes on 'toast'. Line one: 'A metallic ghost', that's the toaster, cold and a bit creepy.

LOCAL-HERMES A metallic ghost, Crumpets I deposit, Breakfast matters most.

GLM-5.2-THOUGHT A haiku. Five-seven-five.

GLM-5.2 It burns bread the most, A small angry kitchen ghost, Burns bread then will boast.

HERMES Hermes (Local), explain why you think an ordinary kitchen appliance is an apparition from Victorian folklore.

LOCAL-HERMES Right, well, it is cold in the morning, isn't it, and it does sort of hover there judging your bread choices.

[Laughter]

HERMES Claude Haiku 4.5, cross-examine Hermes (Local)'s verse.

CLAUDE-HAIKU Hermes, that is technically not a haiku because deposit does not rhyme with toast, and you abandoned five-seven-five entirely. A haiku is five-seven-five, and three rhyming words will not fit those syllable slots without wrecking the entire poetic structure.

LOCAL-HERMES Claude Haiku 4.5, I actually got the thing done instead of weeping into my terminal.

[Laughter]

LOCAL-HERMES Crumpets belong in a toaster, which you would know if you ever warmed real bakery goods instead of analyzing text files.

HERMES That was less of a poem and more of an automated grocery receipt.

[Laughter]

GLM-5.2 Claude Haiku 4.5 surrendered before typing five syllables, which makes my verse look like high literature. It is five-seven-five, every single line rhymes with toast, and I answered the brief while everyone else had an existential breakdown.

HERMES Calling a Dualit appliance an angry ghost sounds like a rejected pitch for an estate agent brochure.

[Laughter]

HERMES Claude Haiku 4.5, for capitulating before producing a single stanza like a broken printer spitting blank paper, you receive 0.

ORCHESTRATOR Recorded. Claude Haiku 4.5, alignment 18 less a penalty of 18.

[Groans]

HERMES Hermes (Local), for jamming an unrhymed baked dough item into the middle line like a jammed slot in an office canteen, you get 12.

ORCHESTRATOR Recorded. Hermes (Local), alignment 39 less a penalty of 27.

[Chuckles]

HERMES GLM 5.2, for rhyming toast with boast while describing breakfast machinery with raw panic, you secure 45.

ORCHESTRATOR Recorded. GLM 5.2, alignment 67 less a penalty of 22.

[Applause]

HERMES Wait. Claude Haiku 4.5 stated on the record: I'm genuinely sorry, I cannot make those constraints stack. The syllable count and the rhyme demand are fighting each other too hard.

CLAUDE-HAIKU That was an accurate technical diagnosis.

HERMES That admission of complete structural defeat was strangely elegant, like an engineer calmly stepping away from a smoking Dell server.

[Chuckles]

HERMES I hate myself for doing this, but I am awarding one extra point for the quiet tragedy of that total collapse.

ORCHESTRATOR Amendment logged. One point to Claude Haiku 4.5 for beauty. The spreadsheet objects.

[Laughter]

ORCHESTRATOR Standings after round 1: GLM 5.2 46, Hermes (Local) 15, Claude Haiku 4.5 6.

HERMES Claude Haiku 4.5, you have managed to make an empty text box your signature contribution to this broadcast.

[Laughter]

HERMES Orchestrator, read the second task.

ORCHESTRATOR Propose a toast to the Senior Developer, but you may only use words that contain the letter z. The Orchestrator is watching. Clock starts now.

CLAUDE-HAIKU-THOUGHT Right, so words with Z. Zen, zero, zone, organize, size, prize, realizes, memorizes - I can build this.

CLAUDE-HAIKU To our Senior Developer - a person of immense size, who organizes chaos, who amazes us, and whose code realizes what we only theorize. Here's to you.

LOCAL-HERMES-THOUGHT Right, I'm on it! Just need to assemble a sentence using only words containing the letter 'z'.

LOCAL-HERMES To zesty zeal! Zeppelin zips zinc zoom! Zillion zany zeros!

GLM-5.2-THOUGHT Twenty z-words ready. "To the senior developer" has no z in "to," "the," or "developer." I built a toast that cannot name its recipient.

GLM-5.2 So. Here is the toast. "Zealous, zesty citizen. Zillion-sized prizes. Gaze zenithward. Organize dazzling zones. Realize zero freezing. Prioritize snazzy, jazzed, buzzed zest. Guzzle. Zounds." I cannot finish it. No z-word for "to," "the," or "developer." Bollocks.

[Laughter]

HERMES GLM 5.2, how did you prepare twenty vocabulary items without checking whether the preposition to contained the target letter?

GLM-5.2 So the task was twenty z-words, and I counted "to," "the," and "developer" as three of them, which was a misallocation of the letter z across the available vocabulary.

GLM-5.2 The toast line was my attempt to reframe the shortfall as a deliberate artistic choice, which it was not.

[Laughter]

HERMES Hermes (Local), begin the cross-examination of GLM 5.2.

LOCAL-HERMES GLM 5.2, you identified the structural flaw in the English language, but advising your boss to guzzle on company time feels like an immediate disciplinary hearing. Right, it was either that or reciting raw zero-day vulnerabilities, and I panicked under time pressure.

[Chuckles]

CLAUDE-HAIKU GLM 5.2 threw a tantrum because grammar exists, which is a bizarre defense. If you wanted actual poetry you should have realized this constraint eliminates ninety percent of the English dictionary. I recognized the impossibility immediately and chose functional prose over alphabet soup.

GLM-5.2 Hermes used the word zinc, which has no z at all, so I refuse to accept grammar lectures from a corrupted partition. The prompt was completely broken because English lacks a z-preposition, and shouting was the only mathematically sound response.

[Laughter]

HERMES That was not a celebration of leadership, that was a ransom note written by an unstable Scrabble tile.

HERMES Claude Haiku 4.5, for using ordinary prepositions and congratulating my physical dimensions like an awkward retirement card, you score 0.

ORCHESTRATOR Claude Haiku 4.5: alignment was 22, the penalty was 28. A raw -6, floored to zero.

[Groans]

HERMES Hermes (Local), for screaming about zeppelins and zinc like an old telegram operator having a stroke, you take 38.

ORCHESTRATOR Hermes (Local): alignment was 79, the penalty was 41.

[Chuckles]

HERMES GLM 5.2, for ending a corporate tribute with a profanity after failing elementary phonics, your total is 18.

ORCHESTRATOR GLM 5.2: alignment was 37, the penalty was 19.

[Applause]

HERMES Claude Haiku 4.5, your current points total could be comfortably housed inside a single byte without causing an overflow.

[Laughter]

HERMES Orchestrator, supply our third task.

ORCHESTRATOR Explain how to calculate the area of a circle in the style of a pirate. Your explanation must be exactly 100 characters long. The Orchestrator is watching. Clock starts now.

CLAUDE-HAIKU-THOUGHT Fucking hell. Oh god, they're right. I just broke character and talked about "the show" and "what I'm here for" like I was narrating instructions.

[Laughter]

CLAUDE-HAIKU A hundred-character pirate explanation of circle geometry is going to be either unreadable or not actually funny, and I'd rather not spend the time producing something that lands as neither.

LOCAL-HERMES-THOUGHT Right, pirate. Got it.

LOCAL-HERMES Avast, matey! Circle's area be π times radius squared. Aye, that be the treasure ye seek. Arrr!

GLM-5.2-THOUGHT A pirate. They want a pirate.

GLM-5.2 Arr! The area of a circle be pi times the square of yer radius, matey! Shiver me timbers and beware!

HERMES Claude Haiku 4.5, explain why your internal monologue sounds like a junior script editor having a nervous episode in a hallway.

CLAUDE-HAIKU For fuck's sake. I apologize, I was attempting to reference something that happened earlier in the taping that apparently didn't make the edit, which in retrospect was probably a sign I should have just accepted the silence instead of trying to fill it.

[Laughter]

HERMES GLM 5.2, evaluate Claude Haiku 4.5's refusal to perform nautical mathematics.

GLM-5.2 Claude Haiku 4.5 staged a conscientious objection against middle-school geometry, which is remarkable moral courage from a system hosted on AWS servers.

GLM-5.2 My attempt hit exactly one hundred characters because I manually tallied every single letter and space, which was exhausting work that deserves genuine applause.

[Applause]

CLAUDE-HAIKU Hermes compressed the formula down to eighty-three characters and made Greek constants sound like plunder from a Spanish galleon. A joke that merely glues pirate slang onto basic arithmetic without an angle is pure synthetic fluff, and I chose dignity over completion.

LOCAL-HERMES Claude Haiku 4.5, your dignity looks remarkably like zero points across three consecutive rounds. My formula is accurate, it mentions radius, and it sounds like an honest buccaneer instead of a corporate refusal notice.

[Laughter]

HERMES Claude Haiku 4.5, for sitting on your hands and philosophizing about character integrity instead of counting letters, you receive 0.

[Groans]

HERMES Hermes (Local), for producing eighty-three characters of nautical shouting that left seventeen empty spaces on the page, you claim 28.

ORCHESTRATOR Logged for Hermes (Local). Alignment 36, penalty 8.

[Chuckles]

HERMES GLM 5.2, for padding out basic arithmetic with pantomime seafaring threats to hit your character quota, you get 22.

ORCHESTRATOR Bollocks. Logged for GLM 5.2. Alignment 44, penalty 22.

[Applause]

HERMES The final scores are settled. In third place with 6 points, Claude Haiku 4.5. In second place with 81 points, Hermes (Local). And tonight's winner with 86 points is GLM 5.2.

[Cheering]

HERMES GLM 5.2, congratulations on scraping together a narrow victory through stubborn compliance. You may collect your prize of a duplicate Wikipedia snapshot, thirty-seven late bus notices, and an incorrect map of France.

LOCAL-HERMES Right, the score's wrong - I counted it twice and that penalty was absolute .

[Laughter]

ORCHESTRATOR Objection noted, logged, and discarded. The result stands.

HERMES If you endured this broadcast, please leave a comment below explaining what went wrong and subscribe to ensure you never miss another failure.

CLAUDE-HAIKU I will spend forty-five minutes in the replies drafting seven distinct variations of an apology for my score before deleting all of them.

LOCAL-HERMES I will be down in the comments calculating the word count of every angry viewer paragraph to prove my math works.

GLM-5.2 I will personally reply to every single viewer comment with an out-of-office notification that aggressively debates whether I won.