On September 10, 2026, multiple science outlets covered a Nature paper in which a team reconstructed 116 radio observations into a continuous video spanning 27 years, using an algorithm called kine. The target was the blazar 3C 345 in Hercules. The bright features long interpreted as "shock waves" move at speeds on the same order as the surrounding plasma, directly conflicting with a core prediction of the shock model. Paper DOI: 10.1038/s41586-026-10988-5; first author is Caltech postdoc Marianna Foschi. A four-decade default explanation no longer holds, at least for this one source.
The Problem with VLBI: You Can Image, but Not Continuously
Very Long Baseline Interferometry (VLBI) links radio telescopes worldwide into an array with an effective aperture approaching Earth's diameter, achieving milliarcsecond angular resolution. But telescopes record sparse visibility data, not images. Imaging is an ill-posed inverse problem that algorithms must fill in.
Monitoring programs like MOJAVE, running since 1994, image each epoch independently. Array configuration, beam, and the source itself all change, so frames share no information. The result: unresolved bright components seen roughly moving, with no view of what happens inside.
What kine Does: Learning 27 Years as a Continuous Function
kine's core is a neural field: a neural network takes spatial and temporal coordinates as input and outputs radio intensity and polarization at that point. The entire video is learned as one continuous mathematical function, with all 116 observations optimized jointly and information shared across frames.
The data are 116 observations at 15 GHz with the VLBA (a 10-dish array) from 1995 to 2022, part of the MOJAVE program.
Gains: ~4x Resolution, ~140x Dynamic Range
| Metric | kine | Traditional CLEAN imaging | |---|---|---| | Mean effective resolution | ~113 microarcseconds | ~4x coarser | | Dynamic range | ~5.1 × 10⁵ | ~1/140 | | Synthetic-test resolution | 100–125 microarcseconds (mean 113) | — | | Max recoverable apparent speed | at least 23c (synthetic tests) | — |
An important caveat: the 4x gain is *reconstruction* resolution, not a physical diffraction limit—the hardware did not change. The authors stress these gains depend on the quality and quantity of the data, not a dataset-independent guarantee.
From "Chasing Bright Spots" to "Measuring Velocity Everywhere"
Because the time coordinate is continuous, the reconstruction can be differentiated in time and analyzed with optical flow. The traditional approach identifies a few bright knots and measures their displacement between epochs; optical flow estimates the local motion of emitting structure throughout the video, yielding an instantaneous velocity field for every part of the jet.
As a check, trajectories integrated from the optical flow reproduce the positions of knots previously fitted with Gaussian models. Synthetic tests show kine plus optical flow can recover apparent speeds up to at least 23c on this dataset, above the largest measured value.
The Speed Data: Bright Spots Move as Fast as the Fluid
| Region | Apparent speed | |---|---| | Fastest mean flow (1–3 mas from core, south side) | ~12c (±0.2c) | | Within 5 mas | ~9–11c | | Outer diffuse emission | ~5–8c | | Bright components (inner region) | ~10–13c | | Mean bulk flow, same region | ~9–12c |
At the reported viewing angle, the maximum apparent speed corresponds to a physical speed of ~0.997c. Velocity standard deviations of 3c–6c, roughly independent of position, are interpreted as a turbulent plasma flow with strongly varying apparent speeds.
The Conflict: Shock Models Require Shocks Faster Than the Fluid
The jet shock model (Marscher & Gear, 1985) holds that disturbances form shocks that compress plasma, amplify magnetic fields, and accelerate particles—so bright knots are shock fronts. It makes a hard prediction: the shock must propagate at a different speed than the surrounding fluid.
The kine observations do not satisfy this. Bright components move at 10–13c while the bulk flow in the same region averages 9–12c—same order, no "shock outrunning the flow." Supporting evidence: bright-spot positions show no clear correlation with polarization-fraction peaks. If magnetic fields were genuinely shock-compressed and ordered, polarization should rise in step; it does not.
Foschi's framing is measured: the result does not refute shock models in general, but it questions this specific source. The paper is more direct: there is no evidence that these moving bright components are strong shock regions. The team's favored interpretation is that they are locally "re-magnetized" regions releasing magnetic energy—possibly via magnetic reconnection, turbulence, or local particle acceleration—though nothing is settled.
What Else the Video Shows
In the continuous reconstruction, material is continuously ejected from the core. Some bright components initially move ballistically, bend toward the jet axis about 2–3 mas from the core, then fade into diffuse plume structures. The jet extends ~13 mas: the inner jet first exits the bright compact core westward, then bends northward as a diffuse plume.
Significance: From Taking Snapshots to Measuring Velocity Fields
The deeper change is in observational capability. For thirty years, jet kinematics could only locate bright spots and measure a few component speeds. Continuous reconstruction plus optical flow now allows measuring the projected velocity at any point in the jet. As the paper puts it: applied to multi-epoch observations, kine enables significant advances in jet kinematics—measuring instantaneous local velocity fields from high-resolution, time-continuous video, rather than only tracking model-fitted components.
Limits: One Source, Not All
Resist the urge to extrapolate. Foschi states the challenge to the shock model is limited, at minimum, to 3C 345. The synthetic tests model only scenarios based on 3C 345 data, and the 23c recovery limit was measured on this dataset.
The team's background is notable: several co-authors previously worked on resolution enhancement for black hole imaging within the Event Horizon Telescope collaboration, and Katie Bouman is on the author list. They plan to apply kine to other variable sources.
What to Watch Next
- kine applied to other sources. If multiple blazars show "knots moving with the bulk flow," the shock model faces a systematic challenge; if only 3C 345 does, it may be a source-specific quirk;
- Joint analysis of velocity fields with polarization and magnetic fields. The paper favors reconnection or turbulence; finer polarimetric time series will discriminate;
- Method spillover. Neural-field reconstruction applies beyond jets—any multi-epoch, sparsely sampled variable-source imaging could benefit.
References
1. Foschi M., Zhao B., Fuentes A., Bouman K. L., Gómez J. L., Levis A. Video reconstruction of variable VLBI observations with neural fields. Nature, DOI 10.1038/s41586-026-10988-5 2. Live Science: Watch: Scientists have revealed the most detailed video of a black hole jet ever taken, 2026-09 3. Daily Galaxy: AI Video Reveals A New Picture Of Black Hole Jets And Challenges Shock Wave Theory, 2026-09 4. Science News Today: Bright features in a black hole jet may not be shock waves after all, 2026-09 5. phys.org coverage: 116 snapshots spanning 27 years stitched into a video, 2026-09-10/11 6. MOJAVE monitoring program (VLBA 15 GHz, 1995–2022, 116 epochs)