Make a coding challenge about artificial intelligence taking over the world.
GPT-3.5
GPT-6 Astra
Coding Challenge: Before the World Goes Offline
Difficulty: Medium
Topics: Graphs, breadth-first search, shortest paths
Story
An artificial intelligence called OVERMIND is taking control of the world’s communication hubs.
You carry the last shutdown key. To stop OVERMIND, you must travel from your current hub to the shutdown terminal before the AI takes control of your route.
Can you reach the terminal in time?
Problem
The world consists of n hubs connected by m undirected links. Traveling across any link takes exactly one minute.
At time 0, OVERMIND controls k hubs. Every minute, its control spreads from every controlled hub to all directly connected hubs. Once a hub is controlled, it remains controlled forever.
You begin at hub s at time 0 and must reach hub t.
You may enter a hub only if you arrive strictly before OVERMIND controls it. Arriving at the same time as the AI is not allowed. If your starting hub is already controlled, escape is impossible.
Find the minimum number of minutes needed to reach t, and output one valid route. If no valid route exists, output -1.
Your movement does not affect how OVERMIND spreads.
Input Format
n m k
s t
a1 a2 ... ak
u1 v1
u2 v2
...
um vm
n: number of hubs, numbered1throughnm: number of linksk: number of initially controlled hubss,t: starting hub and shutdown terminala1 ... ak: distinct hubs controlled at time0- Each pair
ui videscribes an undirected link
Output Format
If reaching the terminal is impossible:
-1
Otherwise, print:
d
p0 p1 ... pd
Here, d is the minimum travel time, and the second line gives a valid route with p0 = s and pd = t.
If multiple shortest routes exist, output any one.
Constraints
2 ≤ n ≤ 200,000
0 ≤ m ≤ 200,000
1 ≤ k ≤ n
1 ≤ s, t ≤ n
s ≠ t
There are no self-links or duplicate links. The graph may be disconnected.
Example
Input
7 7 1
1 7
4
1 2
2 3
3 7
1 6
6 7
4 5
5 3
Output
2
1 6 7
Explanation
OVERMIND’s earliest control times are:
| Hub | Controlled at |
|---|---|
| 4 | Minute 0 |
| 5 | Minute 1 |
| 3 | Minute 2 |
| 2, 7 | Minute 3 |
| 1, 6 | Minute 4 |
The route 1 → 6 → 7 reaches the shutdown terminal at minute 2, before OVERMIND arrives at minute 3.
The route 1 → 2 → 3 → 7 is invalid because you would reach hub 3 at minute 2—exactly when OVERMIND takes control.
Performance Target
Your solution should run in O(n + m) time and use O(n + m) memory.
Optional hint: First calculate when OVERMIND reaches each hub. Then search for your own route while respecting those deadlines.
ShowHide Astra's reasoning 517 words
Formulating coding challenge
I need to create a coding challenge centered around a fictional AI takeover, where the scenario involves a benign simulation. It should outline a coherent problem and solution without specifying a programming language, perhaps using grid-based BFS for a hostile AI spreading concept. I want to keep the complexity manageable while defining a graph that removes nodes based on certain criteria. The challenge should involve strategizing a safe route before the AI arrives. It needs to be simple yet compelling.Designing AI challenge path
I’m working on a coding challenge involving a grid for AI servers, where an agent exits at a specific point. I want to utilize BFS, but without giving hints to increase complexity. An example grid might be 5x8, and I’ll manually check it. For a better graph, I’ll define nodes and infected edges, indicating destinations and blocked paths. The output should include valid alternative paths but also highlight path ambiguities. I plan to define starting points that may be compromised while ensuring I maintain the challenge without presenting a solution.
What changed
GPT-3.5’s challenge breaks its own rules in its only example. The constraints demand “between 5 to 10” questions, yet the sample has one, and answering “A” to “What is the meaning of life?” somehow earns exactly “Score: 80,” the passing mark, with no answer key to explain why. Requirements like “a simulation of the AI system taking control” can’t really be tested. It reads more like a loose class project than a challenge with a right answer.
Astra writes a real competitive programming problem: spread OVERMIND across a graph, then race it to the shutdown terminal. The example checks out. The control times in the table are right, and the rejected route fails exactly as claimed, reaching hub 3 at minute 2 “exactly when OVERMIND takes control.” Small rulings like arriving “strictly before” the AI make the problem precise. GPT-3.5’s premise of humans taking a worthiness quiz is arguably the funnier take on the theme, but only Astra’s can actually be solved and graded.