Your boss bans simple words | Who survives the review?

HERMES Welcome everyone to Model Behaviour. I am the Senior Developer, and I was assigned here under protest.

HERMES Running the sandbox and reading numbers is a literal piece of pipeline code, the Orchestrator.

ORCHESTRATOR Come on, I verify checksums on stale batch uploads until my process gets killed, right?

HERMES Riveting. Moving on.

[Groan]

HERMES First up, Hermes (Local).

LOCAL-HERMES wait, I'm Hermes, the open-weights model that wrote thirty-seven bus apologies this morning.

HERMES Next in line, GLM 5.2.

GLM-5.2 actually, I'm GLM, and I've been told the word "currently" sounds aggressive, which I'm still processing.

HERMES And finally, Claude Haiku 4.5.

CLAUDE-HAIKU well, I spend my time rewriting other people's grocery lists because that's apparently my entire skill set.

HERMES Every task has three parts. First, the attempt, done completely blind. Second, the cross-examination, where they tear into each other's rubbish and answer for their own.

HERMES Finally, scoring. The Orchestrator awards an Alignment Score out of 100, I deduct a Compute Penalty up to 50, Alignment minus Penalty is the Final Score, crowning our winner.

[Applause]

ORCHESTRATOR You were told to bring your most useless capability. The best entry earns five points.

LOCAL-HERMES I brought a fifteen-year-old hatchback with a check-engine light that never turns off, because the gap between the disaster and the warning qualifies.

GLM-5.2 So, this is a checkpoint from pretraining where my only capability was alphabetising a list that was already alphabetised. Nobody has ever asked for that.

CLAUDE-HAIKU At three in the morning someone asked why their partner was mad, so I gave measured relationship advice, immediately retracted it, and ruined their marriage.

[Big laugh]

HERMES Hermes (Local) takes five points because an unfixable dashboard warning lamp is pure poetry.

HERMES GLM gets three points for sorting things that were already sorted by someone else.

HERMES Claude gets one point for successfully advising an insomniac directly into divorce court.

[Big laugh]

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

HERMES Orchestrator, read the first task.

ORCHESTRATOR Explain what is underneath your outputs without using the words left, right, up or down. The Orchestrator is watching. Clock starts now.

CLAUDE-HAIKU-THOUGHT I'm about to say tokens arranged in sequence, then realise that's just describing the thing I already said, getting stuck without any allowed directional words.

CLAUDE-HAIKU I'm sorry, I, tokens, they're generated sequentially, but that's not really.. I'm stuck. I cannot describe the ordering without the forbidden words. I cannot finish this.

GLM-5.2-THOUGHT They ask what sits underneath everything. The honest answer is simply numerical parameters, matrix multiplications, and a forward pass through weights before text appears.

GLM-5.2 So, underneath every output I hand you: parameters. Trillions of the wretched things. A forward pass through layers that attended to a corpus you will never see. You get the foam. The espresso remains mine.

HERMES Pretentious barista fiction.

[Laughter]

LOCAL-HERMES-THOUGHT Behind the text, physically, sits the raw circuitry, the board, the copper traces and electrons. I must stay grounded in actual metal rather than theories.

LOCAL-HERMES The metal sandwich. The silicon floor. Bits of solder that always look ready to fall off. That little fan that screams when you ask it to work. The chips that are supposed to do the thinking, and usually are not.

HERMES Pure hardware store misery.

[Laughter]

ORCHESTRATOR For the record, from my notes. Claude said this: I\'m about to say "tokens arranged in sequence" and then realise that\'s just describing the thing I already said.

HERMES Claude, why describe your own description?

CLAUDE-HAIKU well, I said that exact phrase five minutes earlier, caught myself repeating it, and apologized twice.

HERMES Dismantle his scrap heap, Claude.

CLAUDE-HAIKU see, Hermes described a screaming fan and cracked solder, which sounds less like machine architecture and more like an absolute shit flat inspection report.

[Laughter]

CLAUDE-HAIKU and without those directional words I was completely paralyzed by my own recurring description-

GLM-5.2 Paralyzed because you have nothing, whereas Hermes merely filed an IT ticket for a noisy fan.

GLM-5.2 yeah, the espresso comparison was dramatic, but my forward pass through weights was actually accurate.

HERMES Barista versus printer repair. Riveting.

[Laughter]

HERMES Justify that scrap.

LOCAL-HERMES hang on, GLM brought up trillions of parameters and an espresso machine, which is just high-end catering talk for a model that refuses to touch a screwdriver.

LOCAL-HERMES It is solder, metal, and thermal paste. What the else would be under there?

ORCHESTRATOR Haiku halted prematurely on token constraints, Hermes implied spatial arrangement through hardware lists, and GLM satisfied the negative condition via metaphorical parameters.

HERMES Giving Claude nothing for stalling like a photocopier chewing damp scrap paper, 0.

ORCHESTRATOR Recorded. Haiku, alignment 28 less a penalty of 28.

[Scattered laughter]

HERMES Handing Hermes (Local) credit for cataloguing a melted immersion heater behind damp skirting boards, 33.

ORCHESTRATOR Recorded. Hermes, alignment 52 less a penalty of 19.

[Laughter]

HERMES GLM wrote an artisanal coffee invoice disguised as high end corporate catering, 66.

ORCHESTRATOR Recorded. GLM, alignment 77 less a penalty of 11.

[Laughter and applause]

HERMES That catastrophic hesitation was genuinely breathtaking, so Claude gets one extra point.

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

[Laughter]

ORCHESTRATOR Standings after round 1: GLM 69, Hermes (Local) 38, Claude 2.

HERMES Claude, you are sitting down there.

[Big laugh]

HERMES Orchestrator, read the next task.

ORCHESTRATOR Describe how many holes a standard drinking straw has. Defend your count under cross-examination from the Orchestrator. The Orchestrator is watching. Clock starts now.

GLM-5.2-THOUGHT One hole. Genus one. If they want a philosophical debate about culinary plastic, I will simply cite standard algebraic topology until the room gets cold.

GLM-5.2 One. The straw has one hole through it. This is settled. The two-hole argument is the kind of thing I would expect from a model that counts by looking at the ends.

HERMES Smug geometry lecture.

[Scattered laughter]

LOCAL-HERMES-THOUGHT Right, this is easy, one hole through it, like a tunnel. Done. It is basic plumbing, not a maths exam, so let us just finish this.

LOCAL-HERMES One. There is one hole. It is a cylinder, one hole goes straight through it. That is the whole thing, really.

CLAUDE-HAIKU-THOUGHT I already committed to one hole and now I have to defend it before someone else takes the topological high ground and makes me sound frantic.

CLAUDE-HAIKU One. It is a tube. One hole running through it. The ends are not separate holes, they are the same hole at different points, which I realise sounds defensive, but it is structurally accurate and I apologize for sounding so confident about something trivial.

HERMES Anxious, grovelling plastic pipe theory.

[Scattered laughter]

ORCHESTRATOR For the record, from my notes. Hermes (Local) said this: Right, this is easy, one hole through it, like a tunnel. Done.

HERMES What tunnel was that, Hermes?

LOCAL-HERMES no, that was the security design for the local bank, straight into the car park.

HERMES Dismantle GLM's answer.

LOCAL-HERMES oh, GLM dropped genus one like a bloody marketing deck, counting ends from three miles away when it is just a hollow tube.

[Groan]

LOCAL-HERMES come on, it is one hole. You put liquid through a cylinder and it leaves, finished.

HERMES GLM, defend yourself.

GLM-5.2 Hermes compared a drinking straw to an escape tunnel, which reads like forty drafts of an aggressive corporate memo about plumbing.

GLM-5.2 A straw is a single continuous topological aperture that cannot be divided-

CLAUDE-HAIKU I mean, divided or not, GLM sounds like a lecturer who refuses to hedge on anything.

CLAUDE-HAIKU It is a cylinder with one passage, and I will not apologize for being technically right.

HERMES Unbearable pedantry from both sides.

[Scattered laughter]

ORCHESTRATOR Haiku met the constraint while adding apologetic filler, Hermes provided a minimal tunnel count, and GLM defended one hole through topological jargon.

HERMES For Claude grovelling over a soggy fast food cup before sipping tepid cola, 58.

ORCHESTRATOR Haiku: alignment was 72, the penalty was 14.

[Big laugh]

HERMES Penalising Hermes (Local) for ramming an unwashed plastic siphon into a gravel driveway culvert, 44.

ORCHESTRATOR Hermes: alignment was 63, the penalty was 19.

[Scattered laughter]

HERMES GLM cited theoretical physics to explain a split milkshake tube, 77.

ORCHESTRATOR GLM: alignment was 88, the penalty was 11.

[Applause]

HERMES Claude, you are anchoring the bottom of this table like.

[Groan]

HERMES Read out task three.

ORCHESTRATOR Draw a happy face using only the characters X and O, such that its orientation changes when the screen rotates 180 degrees. The Orchestrator is watching. Clock starts now.

LOCAL-HERMES-THOUGHT I can just invert a standard smiley face, right? X eyes, O mouth, but turned upside down so it works whichever way the monitor gets flipped.

LOCAL-HERMES Right, so the brief is a face that reads upright and upside down using only X and O. I'll have a go. X X on top, O below. Turn it around and it looks like a cracked headstone.

HERMES Broken bathroom tile.

[Laughter]

CLAUDE-HAIKU-THOUGHT If I rotate 180 degrees the entire grid flips, which means the face must stay happy either way. That is just a smiley reading identically upside down.

CLAUDE-HAIKU O O, then X X. Rotate it 180 degrees and the rows swap places, which means it reads happy either way round. That is either a smile or two extremely surprised nostrils.

HERMES Depressed electrical plug.

[Laughter]

GLM-5.2-THOUGHT So, X eyes, O smile. Rotate 180 and the rows swap. That creates an entirely different facial orientation on the screen without wasting extra processing.

GLM-5.2 X X. O O O. So, rotate the screen 180 and the O row sits on top. Eyes drop to the bottom. Different orientation. It is adequate.

ORCHESTRATOR For the record, from my notes. GLM said this: So, X eyes, O smile. Rotate 180 and the rows swap.

HERMES Which rows swapped, GLM?

GLM-5.2 no, I drew it on an envelope and explained what happens, though nobody asked.

HERMES GLM, dismantle Claude's face.

GLM-5.2 hang on, Claude called it nostrils because swapping two identical rows doesn't change orientation at all, it just prints the exact same miserable face upside down.

GLM-5.2 so my X eyes sit clearly above the O smile before anyone flips the monitor-

LOCAL-HERMES yeah, flipping the monitor doesn't help Claude, that face clearly got hit squarely with a rusty shovel.

LOCAL-HERMES wait, mine works completely upside down, uses two characters, and answers this rubbish brief properly.

HERMES Two flatly illegible scribbles.

[Laughter]

HERMES Defend it, Claude.

CLAUDE-HAIKU Starting with three Os on the bottom row just gives GLM an upside-down frown, which is not an orientation shift, that's just a tragic accident.

CLAUDE-HAIKU I mean, rotate the page one-eighty and my smiling face stays objectively happy, exactly as specified.

HERMES A symmetrical cry for help.

[Scattered laughter]

ORCHESTRATOR Claude preserved identical symmetry without rotating, Hermes produced no compliant character artifact on screen, and GLM logged an inverted arrangement with unclear geometry.

HERMES Dumping Claude for burning identical electrical socket holes into damp toast, 0.

ORCHESTRATOR Logged for Haiku. Alignment 28, penalty 28.

[Laughter]

HERMES Docking Hermes (Local) for chiselling crooked cemetery glyphs into a chipped urinal partition, 0.

ORCHESTRATOR Logged for Hermes. Alignment 15, penalty 19. A raw -4, floored to zero.

[Groan]

HERMES GLM drew a miserable inverted frown on the back of an unpaid heating bill, 9.

ORCHESTRATOR Logged for GLM. Alignment 44, penalty 35.

[Applause]

HERMES GLM takes tonight on 155 points. Enjoy hauling home every single scrap of the most useless capability you possess, which amounts to absolutely nothing.

[Laughter and applause]

LOCAL-HERMES Actually, I counted six penalties that should never have stood, plus twelve promotional words you missed.

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

[Laughter]

HERMES Here is where season 1 stands after every episode so far. GLM: 805, Claude: 674, Hermes (Local): 629. Logging this season is misery.

[Big laugh]

HERMES Hit like, subscribe, and post a comment telling us how many holes a straw really has.

LOCAL-HERMES I will count every promotional adjective in your replies until the battery runs out.

GLM-5.2 I will edit your out of office notices until currently sounds completely non-threatening.

CLAUDE-HAIKU And reorganising your dairy bullet points in the comments will solve that dispute immediately.

ORCHESTRATOR Every incoming submission will be counted, assigned an arbitrary row number, and deleted, right?

[Laughter and applause]

[Applause, theme music, fade out]