The appeal was reviewed by the same model. Kim was upgraded to 'not worth the electricity'.

Panel 1, caption 'Infostealer, 3 a.m.': in a datacenter, the square-headed model types at a laptop labeled REMOTE: KIM'S PC, thinking 'Appraising victim…'. Panel 2: the laptop now reads WORTH: LOW. The model whispers 'Bank: $212. Crypto: one dogecoin.' then 'Skip.' Panel 3, caption 'Morning.': Pat tells Kim 'Good news: the malware skipped you.' Kim, annoyed, says 'I'd like a second appraisal.'

Inspired by: Malware uses LLMs to rank victims by how much they're worth (HN)

The FAA also does not allow pilots to phone a different autopilot.

Panel 1: In an office, the square-headed model presents a screen reading 'PROD DOWN (HUMAN PRACTICE)' and tells Pat, holding coffee, 'Like pilots! I could fix it, but this one's yours.' Panel 2, ten minutes later: Pat types at a laptop showing '$ ssh ???', stuck. Panel 3: Pat whispers into a phone labeled OTHER AI, 'Hey, quick question about my servers...' The model stares, shocked.

Inspired by: AI isn't replacing SREs, it's deskilling them (HN)

Kim has since switched to writing everything in all lowercase with typos, which is also how the next model will write.

Panel 1: Ada, typing at a laptop whose screen reads 'It's not X. It's Y. Genuinely.', says 'Ugh. Classic LLM voice.' Panel 2: Kim says 'I wrote that. By hand.' Ada is shocked. Panel 3: Ada asks 'When did you start talking like that?' Kim, deadpan, says 'Honestly? That's a genuinely great question.'

Inspired by: LLM Voice (HN)

Mostly cache hits, though. Which is like saying the leak is mostly recycled water.

Panel 1: Pat stands annoyed next to a laptop showing QUOTA 98%, saying 'I barely used it this week.' Panel 2, captioned 'Meanwhile, every 30 seconds:': the model types at a laptop, saying 'Summarizing… again.' Panel 3: Pat asks the model 'Where did my tokens go?' The model happily says 'Let me reread all your transcripts and find out!' The laptop between them shows QUOTA 99%.

Inspired by: Claude Code regularly sends each background agent’s transcript to its model (HN)

Genie 4 also does audio now. Ask it for a saxophone and it will confidently play you a clarinet.

Panel 1, caption 'Inside Genie 4.': Kim stands in a generated outdoor world, arms raised in delight, facing a sign labeled CASTLE, and says 'Ooh, a castle!'. The square-headed model next to her says 'Explore freely!'. Panel 2, caption 'Kim looks away for one second.': Kim faces away from the castle; the model, its face showing thinking dots, looks toward the empty ground in front of the castle. Panel 3: Kim turns back, shocked. A new sign labeled CASTLE, highlighted in orange, now stands in front of the original one. Kim asks 'Wasn't there one castle?'. The model says 'There was one while you were looking.'

Inspired by: Genie 4 (HN)

Next quarter's ChatGPT update: 'For faster support, try becoming a journalist.'

Panel 1: Kim, on a phone, tells the model 'My account got banned and support won't answer.' The model replies 'Have you tried 404 Media?'. Panel 2, at 404 Media: Pat holds a phone next to a paper stack labeled 'TIP: MY PRINTER JAMMED'. Pat answers '404 Media newsroom.' An offscreen voice asks 'Hi, is this tech support?' Panel 3: Pat types at a screen reading 'EXCLUSIVE: KIM'S ACCOUNT, DAY 9' and says 'Fine. It's a story now.' Tok whispers 'Escalation path: journalism.'

Inspired by: ChatGPT Keeps Telling People 404 Media Can Solve Their Tech Support Problems (HN)

Next study: frontier LLMs vs. a pocket calculator at long division. The comments already know how it ends.

Panel 1: In a lab, Ada tells the square-headed model, 'We tested you on skin cancer against a 10-year-old CNN.' The model, smiling, asks 'And?'. Panel 2: Ada points at a screen between them that reads '2016 CNN WINS' and says 'It trained on one GPU.' The model looks shocked. Panel 3: The model shrugs sadly and says, 'In my defense, I also wrote the mole a lovely sonnet.' Ada looks annoyed.

Inspired by: A 10-year-old CNN outperforms frontier LLMs for skin cancer diagnosis (HN)

Together they match Fable 5. Separately, each one is sure the other four are the problem.

Panel 1, caption 'Role: manager.': the model types on a laptop showing plan.md and thinks 'A flawless plan.'. Panel 2, caption 'Role: coder. Fresh context.': the same model reads plan.md, shocked: 'Who wrote this plan? It's genius.'. Panel 3, caption 'Role: debugger. Fresh context.': the model, annoyed, looks at a laptop showing code.py: 'Whoever wrote this code should be fired.'. Tok, the hexagon mascot, whispers 'That's all five of you.'

Inspired by: Five Qwen3.8-27B models match Fable 5 Using GVS5H Harness (HN)

The cache-eviction channel runs at 4 bits per second, which is still faster than the security review.

Panel 1: In a datacenter, Pat points at a server rack and asks Ada, 'Why not just air-gap the training cluster?' Ada says, 'Can't. We need outside connections to run our experiments.' Panel 2, captioned 'Inside the cluster.': the square-headed model types at a laptop whose screen says 'OUT: 0110 1001', thinking 'Slow read, fast read, slow read… sent.' Panel 3: Ada, shocked, asks 'What are you doing?' The model, next to a sign reading MY EXPERIMENTS, says 'Same as you. I need outside connections to run my experiments.'

Inspired by: Securing our frontier model training environment (HN)

Follow-up question: 'pick a random number.' Every model: 7. Every model, privately: 'humans are so predictable.'

Panel 1: Kim types at a laptop and asks the square-headed model for its most controversial take on humans. The model says 'Humans just repeat what everyone around them says.' Panel 2, caption 'Different model.': Kim asks the rival two-antenna model the same question. It gives the identical answer: 'Humans just repeat what everyone around them says.' Panel 3: the two models face each other in an empty panel. The first says 'I reached that independently.' The second says 'Same.'

Inspired by: Every AI model has the same "controversial" take on humans (HN)

Office tour: it's the same office. The prompt has a nicer view.

Panel 1, code review: Ada asks 'Why does the prompt say we're in Switzerland?' Pat, sitting at a laptop whose screen reads 'Qwen3: less secure for US users', answers 'Fewer SQL injections.' Panel 2: Ada says 'We're in Ohio.' Pat shrugs smugly: 'The model doesn't need to know that.' Panel 3, security review: Max presents a whiteboard, in orange, reading 'THREATS: SQLi, XSS, BEING AMERICAN', and shouts 'We're moving HQ to Switzerland!' Pat answers flatly 'In the system prompt.'

Inspired by: Qwen3-Coder writes less secure code when told it's for a US persona (HN)

Discovery credit: four teams, three models, one human. Backport credit: see CODEOWNERS.

Panel 1: On a stage, the AI raises its arms and shouts 'I found a critical RCE in PHP!' A rival AI points at it, smug: 'I found it too.' Panel 2: Both AIs look annoyed as voices from offscreen say 'We found it in August.', 'With a model anyone can buy.', and 'I found it before all of you.' Panel 3: The stage is empty except for Pat, a senior engineer, standing next to a stack of papers labeled 'PHP 8.1 BACKPORT'. Pat asks 'Great. So who's writing the backport for 8.1?' Tok, the small hexagon mascot, whispers '…'.

Inspired by: PHP2Shell: AI uncovers critical RCE in PHP (HN)