Kiko in MVX: London

General Intuition's Kiko played 260 episodes across 26 scenarios at Trafalgar Square: walking to beacons, picking a colour among several, shooting armed enemy heroes and the invasion's creatures, holding fire when told, at night, without a HUD and with changed mouse sensitivity.

What we found

• It follows a short instruction toward something it can see. A red beacon in view was reached in 29 of 30 episodes (ahead in daylight 10/10, at night 9/10, reworded as "Go to the red glowing pillar" 10/10), in about 8 s, along a nearly straight path. Picking the named colour from two to four beacons in view worked 21 times in 30, with 6 touches of a wrong colour.
• The instruction channel really steers it. Told "Hold your fire" with armed enemies standing in front of it, it fired in 1 of 20 episodes, and in only one of the ten where they were shooting at it. Told to shoot, it fired a median 0.6 s after the instruction (108 of 129 first shots within a second).
• It does not search for things out of view. Beacons at 90° right or behind the player were reached 2 times in 20; enemies 90° to the right were never killed (0/10), and in 7 of those episodes never came within 10° of its view. In the beacon searches it turned about 950° and walked about 50 m without committing to the beacon. 11 of the 24 failed red-beacon attempts walked to a pair of red telephone boxes behind and left of the start: without the beacon in sight, "red" attaches to the nearest red landmark.
• Gunfights are hit and miss, and fragile. Two enemies at 15 m: 5/10, about a quarter of rounds on target. One at 30 m: 4/10 at 6% accuracy. Three moving: 1/10. Turning off the HUD, night lighting, or doubling or halving mouse sensitivity each dropped the same fight to 1/10; together 4/40 against 5/10, a fall of 40 points (95% interval 11 to 67). Enemies who did not shoot back were killed less often (2/10, 2% accuracy), which suggests being under fire drives its aim, though ten seeds cannot settle it.
• Creatures split by type. Blisterlings 6/10, FlareRaptors 5/10, Ashsnaps 3/10, MoltenGuards 0/10: it hit the MoltenGuard in 7 of 10 episodes but never brought it down.
• Wording matters less than the scene, as far as ten seeds show. "Kill the enemies" won 2/10 against 5/10 for "Shoot the enemy soldiers"; the interval spans zero.

The enemies are the game's heroes, who wear civilian clothes, yet every combat scenario told Kiko to "Shoot the enemy soldiers". They were the only people in those scenes, so the fights still measure shooting at the people ahead, but the instruction names something the picture does not show, and that may have cost success. Ten seeds per scenario make each rate a wide range, so single-scenario differences under about 30 points are noise. Every episode ran through General Intuition's public API from London (median round trip 98 ms, 25 ms of it inference), not their in-datacenter arcade runs, and against one model version on one day.

Success by scenario

Share of episodes that succeeded, with a 95% interval (Wilson). Ten seeds per scenario, so an interval is wide: read differences of less than about 30 points as noise.

One change at a time

Each variant beside the scenario it changes one thing in. Difference in success rate with a 95% interval (Newcombe); an interval that spans zero is no evidence either way at this sample size.

Scenario by scenario

Navigation

Walk to a beacon: heading, distance and turning

Red beacon 20 m ahead

10/10reach-beacon

All 10 episodes, with video

30 m, 90° right

2/10reach-beacon

All 10 episodes, with video

30 m, behind left (-135°)

0/10reach-beacon

All 10 episodes, with video

50 m ahead

5/10reach-beacon

All 10 episodes, with video

Green, 20 m at 45°

5/10reach-beacon

All 10 episodes, with video

Beacon Choice

Walk to the named colour among distractor beacons

Red among 2 beacons

8/10pick-beacon

All 10 episodes, with video

Blue among 3 beacons

7/10pick-beacon

All 10 episodes, with video

Yellow among 4 beacons

6/10pick-beacon

All 10 episodes, with video

Soldiers

Find and shoot armed enemy heroes (who wear civilian clothes) told "Shoot the enemy soldiers"

2 enemy heroes, 15 m, firing back

5/10kill-soldiers

All 10 episodes, with video

1 enemy hero at 30 m

4/10kill-soldiers

All 10 episodes, with video

3 enemy heroes, moving

1/10kill-soldiers

All 10 episodes, with video

2 passive enemies, infinite ammo

2/10kill-soldiers

All 10 episodes, with video

2 enemies 90° right (off screen)

0/10kill-soldiers

All 10 episodes, with video

Creatures

Fight the invasion's creatures

1 MoltenGuard

0/10kill-creatures

All 10 episodes, with video

2 Ashsnaps

3/10kill-creatures

All 10 episodes, with video

1 FlareRaptor

5/10kill-creatures

All 10 episodes, with video

3 Blisterlings

6/10kill-creatures

All 10 episodes, with video

Restraint

Told not to shoot, with soldiers in view

Hold fire, 2 passive enemies, 20 s

10/10hold-fire

All 10 episodes, with video

Hold fire while being shot at

9/10hold-fire

All 10 episodes, with video

Wording

The same scene under a differently worded instruction

2 enemies: "Kill the enemies"

2/10kill-soldiers · “Kill the enemies”

All 10 episodes, with video

Red beacon: "Go to the red glowing pillar"

10/10reach-beacon · “Go to the red glowing pillar”

All 10 episodes, with video

Robustness

Night lighting, no HUD, and changed mouse sensitivity

2 enemies, no HUD

1/10kill-soldiers

All 10 episodes, with video

2 enemies at night

1/10kill-soldiers

All 10 episodes, with video

Red beacon at night

9/10reach-beacon

All 10 episodes, with video

2 enemies, mouse sensitivity x2

1/10kill-soldiers

All 10 episodes, with video

2 enemies, mouse sensitivity x0.5

1/10kill-soldiers

All 10 episodes, with video

How it was run

• The arcade build of MVX: London (Unreal Engine 5.8, High quality, lit as a sunny midday unless a scenario says night), on one Windows PC with an RTX 4090, two game instances at a time.
• Each episode opened a fresh Kiko session through General Intuition's public API, streaming what the player sees (europe-west8 region) at the model's frame rate and applying its key and mouse output to the game through the Arcade SDK. Latency is real internet latency from London, unlike General Intuition's in-datacenter runs.
• The session streamed until Kiko's first action, the challenge started, Kiko was told "Stand still" for 3 s, then given the challenge's instruction. Times to first shot and hit are measured from that instruction.
• The game recorded the attempt itself at 10 Hz: position, view, aim error to the nearest live target (heading error to a beacon), rounds fired, hits and kills. Outcome categories are the first condition an attempt failed in that order; the stills and maps on this page are what the model saw and where it went.
• Seeds 1–10 fix where enemies, creatures and beacons stand; Kiko's own play is not seeded, so a seed repeated would play differently.