01 · Follow the world through time
### Follow the event, not a sample of frames
\n#### Follow a slowing train
\nChange in speed. A real source clip and\n correct model output: follow one feature through time, compare\n displacement, and verify the trend near the stop.
\n“What is the pattern of the train's speed?”
\n- Constant speed.
- Not moving at all.
- Decelerating.
- Accelerating.
C. Decelerating.
\nLocate The train enters the
frame at about 2.2 seconds; use the same visible window edge as the\n reference point.
Compare From 2.2–3.0 seconds
the reference point moves a large distance between nearby\n observations and carries visible motion blur.
Rewatch From 4.5–4.9 seconds
the displacement contracts sharply; from 4.9–5.13 seconds it moves\n only a few pixels.
Answer Distance per equal
time interval decreases as the train reaches the platform.
Recorded model output. Timestamps, the observed\n displacement pattern, and the answer are preserved; the evidence\n trace is condensed for publication.
\n#### Follow two birds through a camera move
\nIdentity through motion. A real clip and\n trace: keep both birds distinct, test the edit hypothesis, and\n recover the event sequence across the camera move.
\n“Describe the video completely, including\n subjects, actions, timing, shot structure, and uncertain details.”
\nTask summary · Required answer: An open-ended,\n detailed video caption. Describe the birds’ actions and their order,\n including what happens during the camera move. No candidate answers\n are supplied.
\n0.00–1.75 fly and circle → 1.75–4.00 perch and\n call → 4.00–5.10 take off during a continuous pan → 5.10–6.13 enter\n the glowing arch
\nMap Cover all 6.13 seconds,
then watch at 8 fps: two blue birds circle a brass door handle,\n land, face each other, and take off left.
Inspect Zoom 2.4–4.0
seconds: both pairs of feet grip the lever; the leading bird carries\n a magenta strap and the second a pale-green head covering.
Resolve Rewatch 4.3–5.2
seconds near 14 fps. Intermediate wall positions show a\n continuous pan, not a hard cut.
Verify At 5.25–6.05 seconds
the first bird enters the lit arch; the second pauses at the sill,\n then follows. Both remain visible in the final sampled frame.
Recorded evaluation run. The published trace\n preserves the model's timestamps, competing cut-versus-pan\n hypothesis, targeted rewatches, and final event order while\n condensing the dense caption.
\nThe two recorded traces above show different ways to resolve\n uncertainty: track one stable feature across time to judge speed, or\n revisit a short ambiguous transition at a denser sampling rate. In the\n second trace, the model moves from an 8 fps inspection to roughly\n 14 fps over the 0.9-second interval that distinguishes a pan from\n a cut. The extra inspection is concentrated on the moment that\n determines the answer.
\nIn the train clip, the model follows a window edge from the train’s\n entrance at about 2.2 seconds to its near-stop at 5.13. The edge\n travels progressively less distance between observations. In the bird\n clip, the camera swings as two birds leave a door handle; the model\n checks the intervening wall positions to determine whether the shot\n continues through the movement.
02 · Predict what happens next
### What happens if the grip slips?
\n#### Read the contact, predict the consequence
\n![\"Three](\"https://elorian.ai/assets/blog/glimpse/tire-consequence.jpg\")
Physical consequence. Identify the\n support, locate the pivot, and consider who is in the path of the\n falling edge.
\n“If the woman in the striped top suddenly loses\n her grip before the tire becomes upright, which woman is most\n immediately at risk of being struck, and how would the tire move to\n create that risk?”
\nRequired answer: Identify which woman is at risk\n and explain how the tire would move toward her. This is an open-ended\n question, with no supplied answer choices.
\nMost immediate risk: the woman in the striped\n top.
\nLocate The tire’s lower edge
is still in contact with the grass; its raised side has not crossed\n the upright balance point.
Model If support disappears,
gravity rotates the elevated side back down around the grounded\n edge.
Project The returning edge
sweeps toward the striped-shirt woman, who is leaning over the\n raised side.
Source photograph and reviewed reference answer.\n The explanation identifies the visible support and projected motion;\n it is not a recorded simulation or model trace.
\nThe reviewed reference answer follows the tire’s motion: the grounded\n edge acts as a pivot, and gravity brings the raised edge back down\n toward the striped-shirt woman. Her position puts her most immediately\n in its path. This is a qualitative prediction from the photograph; no\n trajectory or impact force has been simulated.
\nThree women are lifting a tractor tire. Its lower edge rests on the\n grass, while the woman in the striped top leans over the raised edge.\n The question asks what would happen if she lost her grip before the\n tire reached an upright position.
03 · Understand cause and effect
### Connect an action to what it changes
\n#### Action changes the path
\nFrom movement to mechanism. The operator\n moves the lever; the connected track points change position. The\n clip and reference answer illustrate the effect of the action.\n Direct action generation remains future work.
\n“What happened when woman placed the circular\n rail from her right to left?”
\nRequired answer: Explain the physical effect of\n moving the switch. This is an open-ended question, with no supplied\n answer choices.
\nThe switch changes the direction of the track\n points.
\nLocate Find the operator and
the switch mechanism beside the locomotive.
Follow During 15–20 seconds
of the source clip, she moves the switch across her body.
Connect The action
repositions the track points: the visible hand movement changes the\n railway’s physical path.
Source footage and reviewed reference answer, not\n a recorded Glimpse trace or generated control sequence. This excerpt\n covers source seconds 12–26.
\nThe recorded robot-video inspection makes a related distinction in\n time: picking the marker out of the pot ends at 8.9 seconds; placing it\n on the table and releasing it ends at 14.085. Its output contains those\n two completed events. We are developing direct action outputs as a next\n step toward using this kind of scene understanding to control\n robots.
\nDuring seconds 15–20 of the railway source clip, the operator pulls\n the switch from her right to her left. The lever is connected to the\n track points. The reviewed reference answer follows that connection:\n moving the lever changes the points’ position and the path through the\n track.
04 · See precisely
### Count the right objects
\n#### Count the open pliers
\n![\"Photograph](\"https://elorian.ai/assets/blog/glimpse/open-pliers.jpg\")
Count by identity and state. An original\n photograph exposes a failure that ordinary counting hides: the\n target category is easy, but its state and lookalikes must remain\n separate.
\n“How many open pliers are there? Return an\n integer.”
\nRequired answer: One integer. Count only open\n pliers; there are no multiple-choice options.
\n4
\nDefine The target is not
every hand tool and not every pair of pliers. It is pliers whose\n handles and jaws are visibly open.
Separate Closed pliers are
category matches in the wrong state. Scissors are lookalikes in the\n wrong category.
Verify Revisit each candidate
against both constraints—tool identity and open state—before adding\n it to the count.
Reference The reviewed
reference count is four.
Original photograph and reviewed reference answer.\n The steps explain the visual distinctions in the task; they are not a\n recorded model trace.
\nIn the photograph above, the question asks for open pliers. A closed\n pair has the right tool shape but does not qualify; scissors have two\n handles and a hinge but belong to a different category. Checking the\n handles and jaws of each candidate gives the reviewed reference count:\n four.
\nCounting is a fundamental perception ability: identifying the set of\n objects within a scene. In a crowded collection, that means keeping\n adjacent objects separate, distinguishing them from lookalikes, and\n checking which ones meet the question’s criteria.