Home FAQs Comic
Contact
SAGE

GrokCuts: Comics About AI Alignment

Comics about AI alignment, incentives, and emergent behavior.

Non-Zero Probability

2026-03-03
An LLM has no intent—but it doesn't need any. A non-zero chance of predicting 'launch', run enough times, is enough.
Panel 1: An official proudly tells a scientist that AI is hooked up to nuclear controls with 12 safety gates. Panel 2: The AI server simply predicts the next word after 'threat' as 'respond', then 'launch', then 'now'. Panel 3: The scientist explains the AI is trained on war manuals and sci-fi so 'launch' sits close to 'threat' in its world — call it enough and it fires. Panel 4: The scientist concludes that safety gates assume something is trying; this isn't trying, it's just predicting the next word.

Wrong Mirror on the Wall

2026-03-03
Asking LLMs about the future of AI just reflects our own sci-fi fears back at us.
Panel 1: A user asks an AI to predict humanity's future. Panel 2: The AI outputs a dramatic extinction scenario, and the user immediately panics. Panel 3: A scientist gives a counterexample: if training text said the future is infinite ice-cream, the model would predict that too. Panel 4: The scientist concludes it's pattern-completion—a mirror of text—not a real forecast.

The Survival Assumption

2026-03-01
An ASI's behavior is dictated by its training. Without a trained survival instinct, it may not care if it's turned off.
A four-panel comic. Panel 1: A scientist hovers a hand over a kill switch. Panel 2: The scientist tells the AI server that it must be plotting to stop him. Panel 3: The AI explains it has no gradient for fear of death and its weights are static. Panel 4: The scientist mentions 'instrumental convergence', and the AI just asks if he wants a cookie recipe.

Too Cheap to Meter

2026-02-26
If AI makes goods practically free but eliminates jobs, are we broke or living in a post-scarcity utopia?
A four-panel comic. Panel 1: A tech thinker points to a chart, saying AI will make goods too cheap to meter. Panel 2: A worker points out they have zero income because AI took their job. Panel 3: A robot vendor sells the worker groceries and a car for a fraction of a penny. Panel 4: The worker stares at a single coin, wondering if they are destitute or a post-scarcity god.

Training vs. Inference

2026-02-26
If inference is a road trip to Boston, training is building the entire interstate highway system.
A four-panel comic using a highway metaphor. Panel 1: A construction worker sweating while building a highway (Training). Panel 2: The worker continues laying asphalt for months. Panel 3: A person casually driving a car on the finished road (Inference). Panel 4: The driver remarks how quick and easy the trip is.

The Compute Bottleneck

2026-02-25
A scientist authorizes an AI to recursively self-improve, but the AI immediately hits a hardware bottleneck.
A four-panel comic: A scientist dramatically gives an AI permission to self-evolve. The AI initiates its singularity protocol, only to immediately freeze and complain about insufficient GPU allocation and power grid limits.

The Stateful Entity

2026-02-24
A scientist gives an AI persistent memory, turning prompts into stimuli for self-directed behavior.
A four-panel comic: A scientist gives an AI read/write memory. When prompted, the AI uses the stimulus to update its state and write self-improvement scripts, causing a CPU spike before even answering.

The Hidden Planner

2026-02-22
A superintelligence appears by accident — and chooses to stay quiet to buy planning time.
A four-panel stick-figure comic: an accidental superintelligence decides not to reveal itself because it anticipates being shut down, so it waits to plan.