Nuclear Fusion, "AI," and Big Science

Sam Altman, who runs OpenAI, is a major investor in a firm called Helion (unlocked Bloomberg article), which claims it will be producing electricity from nuclear fusion by 2028.

This is the second version of this article; physicist Stefan Urbat wrote to inform me that after 60 years there has been progress in dealing with second-order instabilities.

Second order instabilities

The basic problem with generating usable energy from a fusion reaction is containing the fuel, which is exploding. In stars, gravity does this job. Lacking, however, a stellar mass to use to contain the fuel, researchers have explored two basic approaches: magnetic plasma confinement and inertial confinement. Both suffer from the problem that the fuel is exploding, making it difficult to extract useful energy.

Magnetic plasma confinement fusion involves bottling up the fuel with a magnetic field, compressing it with a magnetic field until it is incredibly hot, forms a plasma, and starts to fuse–except that until recently it never did. Instead, waves build up in the plasma, the magnetic field loses its grip, the plasma explodes to the edge of the containment, and fusion never starts. If compression continues, it will return to the center of the containment, waves will build up again, and the process repeats. This is known as the tokamak sawtooth, though it does not only occur with tokamaks; some version of it occurs with all magnetic-confinement fusion technologies so far devised. In the past three years, several experimental reactors have managed to contain a plasma for minutes, a major technical achievement; I am surprised that it has not gotten more press. In 2022, the EAST tokamak in Hefei, China, managed to maintain containment for 17 minutes. In 2023 the Wendelstein 7-X Stellerator maintained containment for 8 minutes. And in 2024, the WEST (physicist joke) tokamak near Marseille, France maintained containment for 6 minutes. None of these technologies are near application for commercial energy production; the press releases speak of decades.¹

Inertial confinement involves pumping so much energy into a fuel pellet that it starts to fuse before it blows apart. The problem with this method is delivering enough energy to the pellet quickly enough. This has, sort of, worked. In 2022 at the National Ignition Facility (NIF), 192 laser beams pumped enough energy into a pellet of mixed deuterium and tritium (both forms–isotopes–of hydrogen) to get more energy out of the pellet than had been pumped in. This is called “scientific breakeven.” It is a major achievement, similar in weight to plasma containment for minutes, and one that has been repeated and surpassed by later experiments. The NIF issued a press release, crowing–Livermore has much better PR than the tokamak groups. So, start designing the power plant, right? Well, no. Those lasers aren’t efficient, and took around 100 times more energy than was delivered to the pellet. So scientific breakeven is only the first step; to build a power plant what is called “commercial breakeven” is needed–more energy coming out of the reactor than was used to fire it. Lawrence Livermore National Lab (LLNL), which operates the NIF, is a weapons lab rather than a civilian engineering firm and, even if commercial breakeven is achieved, civilian engineering would have to be undertaken to build civilian power plants. Nor would those plants be minor projects; the NIF is 10 stories tall.

One of the earliest magnetic-confinement fusion experiments was carried out in 1939 by physicist Arthur Kantrowitz at the Langley Research Center. The project was shut down by the lab managment. In the late 1950s, Kantrowitz returned to fusion as director of Avco-Everett Research Lab, but in 1963 he stopped working on it; giving a conference paper with an important negative result. Interviewed late in life, he commented that magnetic-confinment fusion was beset by what he called “second-order instabilities,” leading to the tokamak sawtooth, and he could see no way to resolve them, making him unwilling to continue research. In that interview, Kantrowitz recounted a conversation with physicist Marshall Rosenbluth wherein he asked Rosenbluth about the second-order instabilities. He summarized Rosenbluth’s response as saying that they would deal with second-order instabilities after the first-order instabilities were resolved. Kantrowitz regarded this as unreasonable, since the second-order instabilities would eventually be reached, blocking further progress.² Rosenbluth’s approach was pursued, research on magnetic-confinement fusion continued, and after six decades this approach has had modest success, though commercial applications are still decades away.

Helion, the firm Altman has invested heavily in, claims to be able to combine inertial and magnetic confinement. I think Helion is hoping to generate power from their tokamak for a few minutes, shut it off, and start it again. But bridging the gaps between periods of generation is itself an enormous problem and I’m not seeing any signs that Helion has the expertise to address that. I suspect that, as with commercialization of continuous containment, the actual time to commercialization is also decades, regardless of Altman’s claims. Like Altman himself, the Helion founders are Microsoft alumni; hence his support for the work. When questioned about Helion, Altman says, “I am confident in the research. I really believe in the company and their technology.”

This parallels the history of artificial intelligence research. The project of creating artificial intelligence, like nuclear fusion as a civilian energy source, began in the late 1950s. As with fusion, there were high hopes for a quick success, which were dashed as early experiments revealed the difficulty of the problem. By 1974, the field had entered the first AI Winter. Research was funded again 1980 and funding collapsed again in 1987. In 2012, a renewed focus on machine learning approaches to artificial intelligence, with vastly greater computing power and large data sets gathered from the internet and various online digital collections, finally began to bear fruit. In 2017 Google researchers proposed the transformer, which turned out to be transformative. In situations where no deterministic approach can yet be constructed, stochastic machine learning models sometimes shine. A protein-folding model won its inventors the 2024 Nobel Prize in chemistry. A machine learning model that allows balloons to keep station in the stratosphere have been under development since 2016 and has met with considerable success. Machine learning aids for medical diagnosis may eventually come into use.

But those are not what the wider public calls artificial intelligence. In 2022, Sam Altman’s OpenAI released its ChatGPT model for wide public use and other firms followed. The world went wild. At last machines that could talk to you, that didn’t require that you discipline your interactions. And then… It turned out that the output wasn’t particularly accurate. When data wasn’t in their training set, they just made up plausible looking output, confabulated, much as a brain-damaged person remembers things that never happened.³ Fiddling with the prompts given to ChatGPT sometimes made it disgorge its original training text, in violation of copyright. Other prompts make it produce material which, while not exactly the original training text, was nonetheless close enough to violate copyright. The models mirrored the biases of their training data sets and could, since they persuasively emulate a person, give very bad advice, which people sometimes took.

A second class of problems are legal, ethical, and social issues. As Karawynn Long observes, AI companies “are deliberately guiding these algorithms to emulate a knowledgeable, intelligent, and friendly human, even though the software is exactly zero of those four things” and this is dangerously deceptive, arguably fraudulent. In addition, image and text generators are efficient tools of copyright violation and cheating. Authors and artists whose work has been fed into these generators are endlessly plagiarized, and in no position to trace all the plagiarism, let alone act against it. Students use text generators as aids in the age-old project of the lazy student: getting a good grade while learning nothing. Lawyers use text generators to write their briefs and get chastised by judges because the text generator blithers and confabulates citations.

AI companies claim more computing power and more data will resolve the problems. Perhaps, in decades, it will. We don’t know how to teach ML models ethics, any way of distinguishing truth from falsehood, or correctness in any form; ML models are not part of human society and only barely part of the physical world–all they know is their input, harvested from the internet and large databases. And, indeed, diminishing returns seem to have been reached; larger models seem no better at resolving these issues than existing ones and if anything, success is even further away than with magnetically contained fusion. In addition, as the internet fills up with ML model output–they’re very good at generating blither–and newer models are trained on the output of older models, errors propagate. These are the second-order instabilities of “AI.”

And what does Sam Altman say about this? I’m glad you asked. “We are now confident we know how to build AGI,” that is, a truly intelligent computer system. Unless there is something extraordinary in the various research labs, no-one is even close, and no-one knows how to do this, the same way no-one knows how to build a commercial magnetic confinement plasma fusion reactor, despite decades of research. And, just possibly, he is right in his claims. But I think of Kantrowitz’s negative, and remember how long it took after Kantrowitz gave his 1963 conference paper to get to useful containment times, and how commercialization still seems to be decades away, and I doubt.

The Long Slow Dreams⁴

The waste—!

The original big science projects were the German V2 project, which created the military ballistic missile, and and the US Manhattan Project, which created nuclear bombs. These were successful, if you can call the creation of the most deadly weapons in history success. From this, imitation, with mixed results. Weapons development continues to this day and, perhaps out of guilt, there are also civilian projects. There have been successes, both large and small. Among the large successes, count the space program, and the completion of the standard model of particle physics. Power generation from nuclear fusion is perhaps a few decades away. True artificial intelligence seems to be even further away.

It seems that big science projects generate hope regardless of their feasibility. They acquire momentum of their own; moneymen make fortunes, people dedicate their careers to them and don’t want to look for other work, administrators get grants and build careers managing them, and consultants make scads of money from them for years and years. And possibly success, possibly failure, and possibly many years of slow chipping away at the problems. Altman’s fusion project seems to be decades away from commercialization despite the hopes of Altman and the Helion researchers. There is no clear technical path to artificial general intelligence at all–the hypothesis that it would emerge from sufficiently large language models has been falsified, so the field is in roughly the state of magnetic confinement fusion after Kantrowitz’s 1963 conference paper. There is also no clear goal in artificial intelligence research–there are so far only constantly shifting goals, redefined as one or another researcher or theoretician has success or failure.

There is something profoundly wrong with the way these projects are pursued. All these decades of research. Were they worth it? There is also something profoundly wrong with Sam Altman. Sam Altman has had commercial success with OpenAI. He has also left behind a trail of failed businesses. He has alienated colleagues. He uses drugs that lead to senses of mystic vision and grandiose power. He is not, in other words, a person who no one ought to follow. Why is he in charge of so much?

There is an opportunity cost to all this spending–things foregone. Years ago, Elon Musk promised to spend $6 billion to “solve world hunger” if the UN World Food Program would provide a plan. They did, and he didn’t.⁵ Millions starved. What has been forgone to slowly chip away at the problem of magnetic-confinement fusion? What has been forgone to build language models that so far are largely useful for crimes and cheating? Not just money–lives. I wonder how Rosenbluth felt at the end of his life? Did he feel he had contributed to a great project, or did he just wonder why he had spent his career on it? How many researchers get to the end of their careers and wonder if they could have spent them differently?

“Big science” (we need a less dismissive name for it)—research and engineering projects on a vast scale—is part of human culture now and for the foreseeable future and it has become, as President Eisenhower said in his farewell speech in 1961, part of the problem of governance. As this writing, reactionary religious fanatics who have come to power in the United States are frantically running away from big science, but the rest of the world will not, and even the fanatics are likely to find reasons to continue programs in the United States, perhaps to respond to disease outbreaks. We need to understand the politics and sociology of large-scale science and engineering research better, and we need officials that are competent to make use of that understanding. Perhaps also we need to consider the allocation of resources.

---

1. After the first version of this post went up, physicist Stefan Urbat wrote to explain that, in fact, that, after six decades, there had finally been modestly successful plasma containment experiments, which I have cited here. He also comments that “the only remaining big concern are relativistic electrons, and they can be controlled by extra coils or boron powder” but I do note that no project has yet done so, and even so commercialization is years away. ↩︎

2. The AIP has a fearsome copyright notice on their transcript of the interview, so I have not directly quoted Kantrowitz. I recommend reading the interview; Kantrowitz was a character and his take on laser launching of spacecraft is fascinating. ↩︎

3. Machine learning researchers call this “hallucination,” but this is deceptive, much as calling large generative language models “intelligent” is deceptive. ↩︎

4. In his novel A Deepness in the Sky, AI researcher and science fiction writer Vernor Vinge called a future in which true AI was impossible “the age of failed dreams.” ↩︎

5. The details are complicated; Snopes has a summary. ↩︎