🦈 MIT Sent Muscle Cells to the Gym. The Robot Now Swims Like a Lazy Shark.

This week MIT grew a robot's muscles in a dish and made them do cardio, an Austrian company hid an entire cochlear implant under the skin, and NASA found out that space sends the bill to your hips. Meanwhile, rat neurons picked up a consulting gig in AI video, and Stanford taught research papers to answer their own questions, which is more than most corresponding authors manage. Five this week. The robot is the slow one.

RESEARCH

Robots made of living muscle have a weight problem. MIT's Ritu Raman says most of them, her own lab's included, are "bulky" and need millions of cells. So her team went the other way and built a swimmer the size of a stick of gum, powered by a single layer of muscle thinner than a hair. MIT calls it the first very thin, two-dimensional, muscle-powered robot that can go somewhere.

The body is a half-millimeter film of gelatin methacrylate (GelMA, or lab-grade Jell-O) whose two halves act as fins. Each fin wears a coat of muscle cells engineered to twitch when light hits them. Light one fin, and it flaps. Alternate, and the robot steers. The trick in the paper in Advanced Functional Materials was the scaffold. The lab's iris-shaped muscle disk from last year only managed movements Raman calls tiny, since its ultrasoft gel shriveled whenever the cells pulled. A stiffer gel with square-bottomed grooves got the cells to line up and pull as a team.

Then the muscles got a gym routine of flashing lights, and the finished robot swam through a petri-dish maze at about four body lengths a minute. MIT notes that this is cow shark pace, an animal presumably chosen after a long search for a slow enough fish.

It also has roughly the autonomy of a cat chasing a laser pointer, since a researcher steered it with a handheld light. And it isn't the quickest of its kind: OstraBot, another muscle-powered swimmer from this year, hit 15.6 body lengths a minute. What MIT is selling is thinness, which should mean fewer cells and, in theory, a cheaper robot.

For now, it's a strip of gelatin that only moves when someone shines a light on it, which also describes most PhD students by third year.

RESEARCH & NEWS

Cochlear implants are the most widely used neural prosthetics around, and still come with a gadget on the outside of your head that dislikes showers, pillows, and bike helmets. MED-EL CEO Ingeborg Hochmair, who helped build one of the first multichannel implants in the 1970s, has wanted to bury the whole thing ever since, and says microphones and batteries held it up.

On September 29, the Austrian company launched TICI Unico, billed as the first and only commercially available cochlear implant with every component under the skin, CE-marked and Europe-only for now. It weighs 20 grams, runs up to 40 hours on a charge, tops up in about an hour, and is expected to still manage a full day per charge after 15 years. You keep hearing while it charges through a puck on the scalp. Also in the shower, in the pool, and asleep.

The hard part was a microphone that works through your own scalp, and MED-EL isn't saying how it pulled that off. What you can read is a feasibility study of six patients, where speech scores with the buried mic were comparable to the same people wearing an external processor, with fifteen adverse events (one serious, resolved). MED-EL paid for it, and its CEO is an author, so season to taste.

The trade-offs are what you'd expect from sealing electronics inside your head. The battery can't be replaced without revision surgery, the implanted hardware can't be upgraded the way an external processor can, and the clinical trial results are submitted to a journal but not out yet.

Philippe Lefebvre, the Liège surgeon who led that study, expected invisibility to be the selling point. His patients mostly wanted to hear around the clock. Turns out the killer app for a bionic ear is the smoke alarm.

RESEARCH

Astronauts get picked for being offensively healthy, sent somewhere that treats the skeleton as optional, and then returned to a planet with stairs. NASA has known for decades that bones thin out in orbit. What it didn't have was evidence that anything actually snaps later.

Now it has some. A NASA-led study in Mayo Clinic Proceedings went through the fractures US astronauts reported at their annual checkups. Fractures in general were no more common. But hip fractures were, specifically after flights longer than 90 days, compared with shorter flights, with the years before a long mission, and with never flying at all. They also showed up earlier in life than hips usually give out.

Lead investigator Jean Sibonga notes that you can't feel bone loss, so an astronaut who feels recovered goes back to hard training on a skeleton that hasn't. Bones don't send status updates.

Before anyone cancels Mars, count the hips. By Earth.com's read of the paper, there were six hip fractures among 379 US astronauts, and two of those followed a long mission. Two. That's why the team leaned on Bayesian modeling, and why Matthew Drake, who co-wrote the accompanying editorial, warns that "a major limitation is accuracy, as astronaut cohort sample sizes are small and spaceflights are rare".

NASA has started ordering CT scans before and after flight, on top of the usual bone density scans, and the authors want countermeasures that keep skeletons at their pre-flight baseline. With missions longer than six months on the schedule, the timing is not subtle.

Shakira said hips don't lie. Apparently they also hold grudges.

NEWS

The headline says AI models are being built from rat brains. What customers actually get is a software patch, and the neurons never leave the lab.

The Biological Computing Company (TBC, an abbreviation worth remembering), founded by neurosurgeons Alex Ksendzovsky and Jon Pomeraniec, grows neurons from rat brains and human stem cells on chips with up to 4,096 electrodes. It feeds the cultures encoded images and video, records how they respond, and boils the recurring activity patterns down into small software "adapters" for existing AI models. One write-up compares those patterns to traffic in Manhattan, stable in aggregate even though every car does its own thing.

On September 22, TBC announced a collaboration with AWS to bring its first product to market, a text-to-video model it says runs 5x faster at 80% lower inference cost, thanks to a layer that adds less than 0.1% to the underlying model. An earlier demo on a Minecraft-trained world model claimed a twofold gain on "the selected video-quality benchmark" and roughly 4.4 times lower inference cost.

Those numbers come with asterisks. The commercial model's base is unnamed, and TNW notes TBC hasn't published benchmarks and that it couldn't verify the figures. We couldn't find a paper either, and "the selected benchmark" is the kind of phrase that makes Reviewer 2 crack their knuckles. Skeptics will also point out that bolting almost any extra layer onto a model tends to help a little. TBC says its adapters beat size-matched alternatives, in TBC's own comparisons. The independent benchmarks are, fittingly, TBC.

To be fair, "we studied neurons and wrote better software" is a perfectly respectable research program. It's a long way from a rat running your video model, though. The rats, for the record, are not on the cap table.

RESEARCH

Every computational biologist knows the ritual. You read a great methods paper, clone the repo, and lose a week to broken dependencies and a README that stops mid-sentence. Emailing the corresponding author is technically an option, in the way prayer is technically an option.

A Stanford team led by James Zou would like to automate the corresponding author. Their open-source Paper2Agent framework, published in Nature last month, takes a paper and its code, wraps the methods as tools, and tests them against the paper's own outputs, with up to six tries before a tool gets dropped. Out comes a chat agent that runs the method on your data.

The test case was AlphaGenome (yes, the DeepMind model from two editions ago). Paper2Agent turned it into 22 validated tools in about 45 minutes for $14 on a laptop, and the agent scored 98.7% on tutorial-style questions, against 82.7% for Claude handed the raw repository.

IEEE Spectrum reports that three linked paper-agents flagged a little-understood gene, GPR137, as a likely psoriasis gene, while 26 of 100 computational biology papers couldn't be converted. Zou counts the failures as a feature. If your paper can't be turned into an agent, your code and documentation probably had it coming.

The paper is more careful than the pitch. A human picked which of the agents' ten proposed checks to run, and the GPR137 signal only reached significance in stimulated T cells. The AlphaGenome agent also disagreed with the AlphaGenome paper about which gene matters at a cholesterol locus, and the authors' verdict is that it's hard to call.

Zou wants future papers to carry an "agent availability" statement next to data and code. Naturally, the team fed the Paper2Agent paper to Paper2Agent, and it now answers questions about itself. Finally, a corresponding author who replies.

If this week had a theme, it was delegation. MIT handed the motor's job to muscle. MED-EL handed the hardware to your skull. TBC handed R&D to a dish of neurons, and Stanford handed the corresponding author's inbox to a chatbot. The astronauts handed bone maintenance to gravity and then left gravity behind, which is the one handoff that went badly.

Most of it holds up once you read past the headline, as long as you check who's holding the flashlight, who paid for the study, and how many hips are in the dataset. That's the part worth doing yourself.

Which job would you hand to a rat neuron? Hit reply, we read everything. And if you know someone still waiting to hear back from a corresponding author, forward this along.

Keep questioning everything (especially anything tested on "the selected benchmark"),

Prateek & Jere

P.S. Two hips is a small sample, but it's still a bigger n than some of our thesis chapters. And terveiset Suomesta, where the sidewalks are a few weeks from icing over and running the same experiment on the rest of us.