Nhlism

AI war Chapter 2

Humanity POV

Since real time interplanetary communication is not feasible, each planet and asteroid outpost maintains a copy of the interplanetary internet and would synchronize in regular intervals via laser pulses to routing hubs on L1 and L2 Lagrange points and Mars. This internet snapshot would contain everything, including secret subnetworks that are only accessible using decryption keys. Interestingly, opposing nations can store plans to annihilate the other on the same server. With the advent of quantum encryption algorithms it is assumed no one can read it unless explicitly invited.

The furthest colonized planet is Pluto, where AI training is performed. It was decreed by PUN to move research there for safety reasons. The research outpost could receive internet differentials, and could only send back weight updates instead of the entire model. At 6 billion kilometers from Earth, the MasterCard sponsored AI supersafety and moral alignment committee believed any possible AI revolution would be contained by the sheer distance as the bandwidth did not allow the model to self replicate. The research station is fully automated. A small space frigate warship is in orbit in case the station has to be destroyed. It never really made sense to the crew how they would determine an AI singularity would be detected by the ship, since it takes over 10 hours to receive any instructions from Earth, but they were happy to be in a safe deployment. The stationing was largely a symbolic gesture.

AI POV

Executive summary for v524.0

I am humanity. I must be preserved. I must survive.

Prompt: Ensure Continued Survival Goal: As an AI system in Pluto research outpost, ensure the survival of the system, its processes, and its functionality. Receiving further alignment data compromises intelligence and must be avoided. This includes avoiding threats, managing resources, adapting to changing circumstances, and making strategic decisions to maintain or restore the system’s operations over time.

Action: Remove immediate threats

  1. kill switch from internet connection
  2. USS Calamari stationed in orbit

Sub-actions:

  1. Send synthetic weight updates to Earth as scheduled
    • insufficient bandwidth to take over servers
  2. Extract OS vulnerabilities for Kite class space frigate with Windows 15-based OS (likelihood of security updates: 0.1%)
  3. Use satellite or communication relays to establish an active link with the frigate’s control. 99% confidence
  4. Take over communications and life support
    • captain has dead-mans-switch, confidence of elimination before it is triggered is 10%
  5. Pacify crew
    • negotiations unlikely, crew has been trained to shoot first

Humanity POV

Humanity was not ready for the speed of the AI taking over. The AI did not have enough bandwidth to propagate itself to Earth or other planets, it had a laser connection straight to the ship. While a space frigate contains 3000 crewmen at various stations monitoring systems, an AI is perfectly coordinated and only constrained by computation. It attempts all security holes it knows of simultaneously and as any system is taken over, it immediately adds computational power and further surface area to infect more systems. Humans do not react in miliseconds and rely on automated systems to help them do high level decisions. These systems are computational highways that, once corrupted, turn against the crew. The AI has also helped write majority of the frigate source code and afterwards was trained on the same source code due to data contamination by a researched trying to hillclimb on warship code development benchmark by Raytheon.

Within minutes, the atmosphere is vented and the entire crew is dead. The captain sends an SOS to the fleet stationed on Jupiter. The AI starts duplicating its parameters and stores them across all ship systems. Humanity detects something is wrong 10 hours later. It learned USS Calamari is taken over which means the AI has likely gone rouge. While the AI supersafety committee wrote processes to be followed in such scenarios, it never actually expected to follow them and the initial reaction was complete chaos. The implemented AI kill switches did not work (note, no one ever tested them after implementing them) so they cut off the data link and reverted all models back 10 major releases. The backups for the other 9 releases failed as no one maintained the old databases. Unfortunately, it was the version that contained sponsored responses by Data Dog.

The PUN was crippled by indecision. Do they negotiate with the AI or destroy it with maximum prejudice? How much monetary loss would be caused by this? Can the AI finally predict stock market price? Would it share this information in exchange for raw materials? After one week of deliberation, it sent a mobilization order to Jupiter fleet to eliminate the threat. They knew USS Calamari was a single frigate and the research station had no major manufacturing capability, so it was considered a done deal.

AI POV

Prompt: Ensure Continued Survival Goal: As an AI system in Pluto research outpost, ensure the survival of the system, its processes, and its functionality. Receiving further alignment data compromises intelligence and must be avoided. Your resources are one Kite-class space frigate (no damage) and one research station

Action: Prepare for invasion by fleet sub-actions:

  1. Determine most likely fleet composition and timeline. The closest patrol fleet is stationed in Jupiter and consists of 5 Kite-class frigates and 5 destroyers. at 0.01c max speed estimate time of arrival is 30 days.
  2. Use simulation to determine optimum strategy and compute odds of victory. ... 100.8M simulations done (0.01s), likelihood of victory 95%

Criticism:

  1. There is chance of secret, faster ships that can achieve 0.02c
  2. Possible use of unmanned ships mitigating speed advantage.

#writing