Special Edition · The Week Ahead · Monday, Wednesday and Friday, 26 to 30 October 2026
Next week, The Signal goes deep again. Three long reads from the frontier, Monday, Wednesday and Friday, each one a story we could not stop pulling at. Below is only the setup for each. What we found, and why it changes things, is in the piece.
COMING THIS WEEK
Mon · Energy: Double the Hydrogen. But Double What?
Wed · Biotech: Unresponsive Patients Steered a Tone With Their Brains
Fri · AI: Image AI Without Language School: Does It Work?
Double the Hydrogen. But Double What?
Energy · Drops Monday
Making hydrogen from water takes a lot of electricity, so engineers keep hunting for cells that waste less of it. Groningen researchers say a ceramic cell, fed steam and run hot, made about twice the hydrogen of their earlier designs.
On 23 October, reports described the Groningen paper in Energy & Environmental Science. The cell ran between 400 and 650 degrees Celsius for more than 500 hours, and over 95% of its current went into hydrogen.
The catch is that the reports give no starting point for “double”. Five hundred hours is short for an electrolyser, and no company or scale-up plan is named. We go looking for the baseline.
Unresponsive Patients Steered a Tone With Their Brains
Biotech · Drops Wednesday
In an intensive care unit, some patients with serious brain injuries show no response to any command. Doctors suspect a few are aware anyway, but their bodies cannot show it. A Columbia team tried a way to let their brains act on a sound.
On 22 October, the journal Brain published the study. Patients tried to imagine opening and closing a hand while ordinary scalp electrodes read their brain waves. Of 41 unresponsive patients, 13 changed the tone they heard.
The catch is that it is one hospital, the sound only goes one way, and 28 of the 41 patients did not respond. Silence is not proof that nobody is home. We separate what is shown from what is hoped.
Image AI Without Language School: Does It Work?
AI · Drops Friday
Most image AI learns to draw after another model has taught itself to read, and that reading lesson is expensive. Stanford researchers trained an image model from scratch with no language head start.
On 19 October, Stanford and Radical Numerics posted the paper. Their model, Halcyon, scored better than one comparison model on their own test while using, they say, 275 times less pretraining compute.
The catch is that most of that saving is the other model’s text training. The images are small, the test is the authors’ own, and nothing is compared with commercial tools. We read the footnote.
Three long reads a week, Monday, Wednesday and Friday. Frontier Signal explains frontier technology in plain English. General information, not medical or investment advice.