From The New Atlantis, October 7:
We’re worrying about the wrong arms race with China.
A curious feature of the last few years of anxieties about AI has been how they favor some dystopian fears over others. Right now, because of ChatGPT’s disarmingly natural conversation skills, we see the domination of worries about AIs becoming new, alien minds. Will AI models become conscious? Will we lose control of them? Will they outcompete humanity in the struggle for existence? But these fears remain abstract and indistinct.
Meanwhile, a more concrete dystopian future is coming into focus, although it does not receive nearly as much attention: widespread genetic engineering. It is quite possible that the most immediate threat AI poses to humanity is not that its superhuman intelligence will beat us in the game of natural selection, but that it will unleash the power of artificial selection to the advantage of some people over others.
AI systems’ mastery of language may or may not portend a future of superintelligent AI minds, but it already provides a proof of concept for a revolution in gene editing. And though such a revolution promises to unlock transformative medical advancements, it also brings longstanding bioethical dilemmas to the fore: Should people of means be able to hardwire physical or cognitive advantages into their genomes, or their children’s? Where is the line between medical therapy and dehumanizing enhancements?
Just as AI precipitates these morally fraught capabilities, the geopolitical race for AI dominance is upending the historic monopoly that Western nations have had in shaping international bioethical norms. China’s remarkable progress in AI, along with its demonstrated willingness to experiment with genetically enhancing its population, raise the possibility that a totalitarian state with profoundly different ethical standards from our own will have at least equal say in determining the future of genetic engineering.
The window of opportunity for the United States to avert the worst outcomes — and to prevent China from wielding decisive influence over the trajectory of AI in biotechnology — is closing.
Since 1953, when James Watson and Francis Crick discovered life’s genetic code along the spirals of a double helix, the dream of biological engineering has seemed tantalizingly close at hand: If we know how life’s information is stored and structured, perhaps we can rewrite it according to our own designs. But whether you see this prospect as a panacea or a new Pandora’s box, progress toward it has been slow, owing largely to the fact that genetic information is overwhelmingly vast, disorderly, and unwieldy.
Scientists have identified some particular genes that determine or influence particular conditions. But that’s a far cry from understanding genomics writ large, as genetic features are more often the result of constellations of genes working together. In a very real sense, genes are a language — a system to record and transmit information — but one that humans are simply ill-suited to speak. Watson and Crick may have discovered life’s alphabet, and subsequent geneticists may have deciphered the meaning of a variety of words, but a lexicon and grammar have yet to be uncovered.
Enter AI. It is likely to be to genetics what calculus was to physics, providing the tools necessary to harness the full power of biology and making earlier efforts appear primitive by comparison. The large language models that power chatbots like Claude and ChatGPT demonstrate how machine learning techniques can crack entire human languages with great sophistication and minimal oversight. That tremendous feat, in principle, is transferable to cracking the language of genomics in ways that were previously inconceivable. Related techniques have already been used to decipher lost human languages that had long stumped linguists. And AI excels in exactly the sort of ultra-complex, multivariable pattern recognition needed to disentangle the meaning of genetic sequences — which are a non-human language whose combinatorial complexity would have remained impenetrable using conventional methods.
At the time of writing, the most wide-ranging and ambitious project using AI to decipher the language of genetics is the Evo 2 system released by the Arc Institute, Nvidia, and others in February. As a report by Stanford University notes, the system is trained on the genetic information of “all known living species — and a few extinct ones,” comprising a dataset of almost 9 trillion nucleotides, in hopes of parsing the function of DNA in every domain of life. It has already shown promise in predicting which among an individual’s many genetic mutations are most likely to contribute to diseases like cancer, in identifying new relationships between multiple scattered genes, and in writing novel genetic code.
Other examples of AI successfully enabling gene literacy are also cropping up. Chinese researchers have built an AI model that can create 3D images of human faces based only on DNA traces, which the researchers believe will be useful for finding missing children even years after they disappeared, or for criminal investigations. Yale, M.I.T., and Harvard have collaborated in using AI to author synthetic DNA sequences that are able to switch genes on or off according to particular circumstances, a major boon for the specificity possible in gene therapy.
AI techniques are not only reading and writing genomic data: they are also creating the tools needed to make more effective CRISPR gene editors, the leading method to manipulate genetic material. Until recently, CRISPR-based gene editors had largely been derived from existing, naturally-occurring microbes and adapted for use in cells of other organisms, including humans. Even while this method was hailed as a breakthrough in the granularity with which scientists could edit genetic material, it has still been plagued by imprecision and “off-target effects” — that is, unintended genetic edits. By feeding massive amounts of protein data into large language models, biologists are beginning to be able to create bespoke CRISPR gene-editing proteins from scratch that are precisely designed for specialized purposes. This method is both more effective at locating and splicing specific genes with fewer misfires, and has a long runway ahead for continued improvements.
To be clear, there are still considerable hurdles. There remains a tremendous amount of mystery in genetics. And the biological complexity that modulates how genes function is formidable.
Even so, the trajectory is clear. As commentators clamor over whether superintelligence looms, AI is quietly bringing about a much more certain future: one in which humanity is able to “speak” — and write — genes with fluency.
Fraught Capabilities....
....MUCH MORE
