AI as questioner, not ghostwriter, is how it belongs in education
Use an LLM for steps 2 through 6 of the research process and let it only ask questions, never generate ideas — that way students aren't fooled by the illusion of intelligence; photography killed realist painting, and AI will force out the part that is uniquely human.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
AI only asks questions, it doesn't think for the student
The researchers break ‘posing a good research question’ into eight steps: choosing a topic, generating sub-questions, expanding keywords, searching and synthesizing the literature, finding knowledge gaps, forming and iterating a specific question, going back to the literature to validate it, and getting feedback from experts and peers. They use the LLM only for steps 2 through 6, building a tool called the Socratic questioner that has the AI only ask the student questions — ‘Is this narrow enough?’‘Have you thought about that?’— forcing the student to iterate themselves, without generating ideas or doing the work for them. The authors' reasoning: used unstructured, an LLM will hand you weakly justified or outright hallucinated conclusions, because LLMs are good at mimicking the structure of human thought and speech but are not themselves deep thinkers, and they create the illusion of being ‘very smart.’
— Steven NovellaResearch method is passed down by mentors, with huge variance
Cara points out that what actually supports this research training at universities is the library science department, but most students never use it. More importantly, this material is usually taught by a mentor rather than systematically: get a good mentor and you're fine, miss one and you're missing a big chunk — the variance is enormous. The people who get ahead often have an intuition for it, and intuition can't be taught; only process and method can. That is exactly the hole this research tries to fill — operationalizing the process, telling you at each step what the goal is.
— Cara Santa MariaAI slop is the baseline; exceed it or you've done nothing
Steven uses the impact of photography on painting as an analogy: once painting realistically was no longer scarce, painters had to do something else. Likewise, if AI slop is the baseline, then if your output doesn't exceed what AI can do alone, you've added no value to the process. The classroom exercise he imagines: take the same search, do it once with AI prompting and once with the feature turned off, and see whether you can find something the AI can't. He also lays out the disagreement plainly: AI entering education is inevitable; what isn't inevitable is whether it leaves a generation intellectually crippled by dependence on AI, or teaches a generation to collaborate optimally with AI and think more clearly.
— Steven NovellaThe robot can outrun Bolt but can't plug in a cord
The second World Humanoid Robot Games were held in Beijing, with more than 2,000 robots competing in events including running, kickboxing, factory work, and restaurant service. Tiangong Ultra ran the 100 meters in 8.64 seconds, faster than Usain Bolt's 9.58; the same robot took 21.5 seconds for the same distance a year earlier, and Jay thinks the huge change came from the software driving it, not the hardware. But Jay puts the emphasis on the other end: plugging an extension cord into a wall socket requires a person to identify the cord, confirm the connector matches, grasp it, judge its orientation, rotate it, align it, insert it with the right amount of force, know when to stop, and troubleshoot if it won't go in — we do all of this outside conscious awareness, and it is the robot's biggest roadblock.
— Jay NovellaAt the competition: cross the finish line and catch fire
Jay stresses that these are only demonstrations, not stable capabilities: robots were still falling during the competition, one came apart into pieces, and another caught fire after crossing the finish line — literally finished in flames. He uses Moravec's paradox to explain the contrast: things we assume are intellectually hard may be easy for a computer, while perception, balance, and manipulating objects — abilities humans don't have to think about consciously — are extremely hard for machines. On the data side, a Chinese robotics company says existing real-world robot training datasets contain roughly 100,000 hours, and they are trying to collect tens of millions of hours, a path similar to Tesla feeding its system with massive driving hours.
— Jay NovellaThe 99% figure comes from a convenience sample
A 2017 study took a convenience sample of 202 deceased players, about half of them NFL players, and found that 99% of the donated NFL brains had CTE, as did 87% of the non-NFL ones. Cara says the number is scary but not surprising, because donation itself carries selection bias: families only decide to donate a brain when the player showed cognitive decline, dementia, or neuropsychiatric symptoms in life, and by the time they donate they already suspect CTE. So the percentage can't be read as a population prevalence — ‘that's precisely why they donated it to you.’ The new study swaps the denominator for all 1,712 deceased former NFL players and gets a range of 18.5% to 98.7%; Cara says that range is so wide it barely says anything, it just sets a floor.
— CaraAntarctic microbes live on hydrogen in the air
The study analyzed 676 microbial genomes and found that energy-metabolism-related genes are horizontally transferred unusually often, with aerotrophy standing out: microbes combine hydrogen, carbon monoxide, and in some cases methane with oxygen to obtain energy, essentially a very slow burn. If that energy is used to fix atmospheric CO2 into biomass, it becomes ‘atmospheric chemosynthesis.’ In Antarctic soils at minus 20 degrees Celsius, microbes can meet nearly all their basic energy needs from these gases. The byproduct is liquid water — on the driest continent, that amounts to pulling water out of thin air. Bob draws the conclusion: astrobiologists have had the slogan ‘follow the water’ for decades, and this work suggests another slogan that may be more productive is ‘follow the hydrogen.’
— BobFusion power scales with the fourth power of the magnetic field
The key to magnetic confinement is how strong the magnetic field can be, and you must use superconducting magnets, otherwise resistance heats up and burns out the whole device. Fusion power scales with the fourth power of magnetic field strength: double the field and the output from the same plasma volume becomes 16 times greater; triple it and you get 81 times. So making the field as strong as possible isn't an optimization, it's the deciding factor. The new approach uses rare-earth ceramic materials, raising the superconducting temperature from 4 Kelvin to around 60 Kelvin, into the liquid nitrogen range; the real headline is pushing the maximum superconducting field strength from 12 Tesla in the old magnets to 20 Tesla. The company also holds something back — what the wire inside the magnet is made of, they say, is secret sauce and they won't tell anyone. The show explicitly notes this recalls past technologies billed as the ultimate battery that refused to show anything, with the difference that this time the other side says the design, materials, and proof of concept are all done.
The stellarator's biggest old flaw is leaking plasma
The show reviews the company's progress along the way: every time it hits a roadblock it solves it, then the next one appears, but it keeps solving them. The latest design at least solves the stellarator's biggest problem for decades — plasma leakage, the hurdle that couldn't be cleared before. That alone is a huge gain and makes the stellarator route one that must be taken seriously, enough to squeeze into the ranks of mainstream players. But the host says plainly that he remains skeptical, and wants to see it actually produce more electricity than goes in before he believes it, and thinks the 2030s are optimistic — a commercial stellarator before 2035 would surprise him.
In their own words · checked verbatim
LLMs are good at mimicking the structure of human thought and speech, right? But they're not deep thinkers themselves. So they kind of create this illusion that what they're saying is is very intelligent, but it's actually very superficial.
Steven Novella19:19
Another one caught on fire after crossing the finish line. It literally finished in flames.
Jay Novella33:29
So, we can't really read to into this percentage of these brains had CTE. It's like, yeah, that's why they donated them to you.
Cara46:42
So now you have a range of 18.5% to 98.7%. That's a big range. That doesn't really tell us anything, does it?
Cara48:46
So when microbes oxidize hydrogen aerobically, water is created. Now, how cool is that? You're in one of the driest places in the world and you could basically pluck liquid water out of the atmosphere right all around you.
Bob1:09:00
because fusion power scales to the fourth power of the magnetic field that means that if you double the magnet the magnetic power you get 16 times the output fusion power from the same volume of plasma fusion and then if you triple it that's 81
the big thing is that the maximum superconducting magnetic field strength was 20 Tesla. 20 Tesla. It stays super Yeah. stays superconducting at 20 Tesla. The previous type of current the current magnets they used were 12 Tesla.
When anyone's looking for funding, I take everything they say with a grain of salt. Right. And until I, you know, I'm skeptical until they freaking do it. is the bottom line till they do it.
Figures
| Tiangong Ultra 100-meter time | 8.64 seconds | 25:25 |
| Usain Bolt 100-meter world record | 9.58 seconds | 25:25 |
| Same robot's 100-meter time a year earlier | 21.5 seconds | 26:26 |
| CTE rate among donated NFL brains in the 2017 study | 99% | 44:41 |
| CTE rate among donated non-NFL brains in the 2017 study | 87% | 44:41 |
| Donated brains diagnosed with CTE / total donated in the new study | 315 / 338 (93.2%) | 48:46 |
| CTE prevalence range calculated across all 1,712 deceased players | 18.5% – 98.7% | 48:46 |
| Maximum superconducting field strength of the old magnets | 12 Tesla | 1:24:24 |
| Maximum superconducting field strength of the new ceramic material | 20 Tesla | 1:24:24 |
Glossary
- Socratic questioner
- An LLM tool that only asks the student questions and never generates ideas, forcing the student to iterate on the research question themselves.
- Moravec's paradox
- High-level reasoning that humans find hard is easy for machines, while perception and manipulation are extremely hard for machines.
- CTE
- A neurodegenerative disease associated with repeated head impacts, diagnosable only by autopsy.
- aerotrophy
- A metabolism in which microbes combine gases such as hydrogen with oxygen to obtain energy, essentially a very slow burn.
- stellarator
- A fusion device that confines plasma with a three-dimensional helical magnetic field; its complex shape once led to it being abandoned.
- secret sauce
- The magnet wire material the company refuses to disclose, which the show uses to draw a parallel with past technologies that wouldn't show their work.
How to listen
Engineers and investors interested in AI in education, the capability limits of humanoid robots, CTE epidemiology numbers, and fusion magnet routes — especially anyone who wants to see the gap between a demo and a stable capability.
The roughly 50-minute to 1-hour-4-minute transition between the CTE segment and the Antarctic microbes can be fast-forwarded.