This is not a race we can afford to lose
AI may well be a stock investment bubble, but the AI arms race is a hell-for-leather battle where the winner takes it all
With AI development on an exponential curve towards self improving models, the nation that gets “there” first may get access to their adversaries encryption codes, crack their firewalls, hack their software, design the better drones, anti-missile weaponry, cure diseases and own the patents the world wants, or create the bioweapons that sabotage crops, generate fuel, or kill or cripple millions of people. That’s assuming AI doesn’t conquer us first. This is why data centres are appearing everywhere and the “clean energy” revolution got left by the wayside.
As Scott Bessent said today “There is no day after tomorrow if we lose”
“Beating China—there is no day after tomorrow if China wins at this,” Bessent said on Sept. 9 at Breitbart News’s “State of the Economy” event in Washington. “If they were to pull ahead of us on AI, then nothing else matters.”
— Arthur Zhang, The Epoch Times
Bessent was answering questions about the rising opposition to datacenters in the USA and also thousands of hostile attacks on American AI trying to copy and steal capability. China would be very happy with any protests. Youtube physicist Sabine Hossenfelder was approached by someone wanting to pay her to tell us AI will kill us all. (And maybe it will). But there is one main beneficiary of anti AI sentiment in the West and that’s China. Given the existential battle, China would be crazy if they weren’t pumping money into grassroots protests about AI datacentres in the US.
China, though, hardly needs to bother whipping up angst in Australia where our electricity grid has been sold out to the Weather Gods and is cripplingly expensive and underpowered. A few days ago Transgrid, Australia’s largest electricity company wrote to data center management to say the NSW grid was nearly full. There was barely 1.5GW to spare in Sydney. We could have led the world when we had a coal fired grid with spare capacity. Nevermind. The race is here and we’re not in it. China has nearly twice the capacity of coal fired plants that it needs, and we barely have enough.
Last month a team created 16 entirely new viruses with the help of AI
These are not viruses you and I could catch, they only infect a bacteria called E.Coli. But if they wanted to make a virus that could infect humans, it could soon be a lot easier and faster than it would have been back in the prehistoric era of say, 2019. This should make anyone sit up and pay attention.
King et al https://www.science.org/doi/10.1126/science.aec2657
A virus that infects bacteria is called a bacteriophage. It’s a tiny molecular-machine in the order of 200nm in length that replicates. Once inside a hapless E.Coli, it commandeers the cells machinery to produce hundreds of clones of itself. This one pictured here is one of the 16 new types of virions. It’s a ball of interlocking proteins that protect the DNA in the centre. Figure K (right) that looks like a bowl of spaghetti exploding, shows just one of the interlocking shell proteins up close. Those are a particular string of amino acids that always fold in this exact convoluted shape. AI can predict the shape of the spaghetti lattice from the DNA code. The shape is determined by things like disulphide and hydrogen bonds, and electrical charges.
AI can leapfrog over thousands of generations of trial and error evolution
Before Large Language Models, a research team had to culture real viruses in the laboratory, and they had to do it again and again to generate random mutations. One experiment ran for 13,000 generations and took six months. That team found the virus had accumulated 11 new mutations every 20 days. Then they’d have to test each mutant to see if it behaved the way they hoped, but it was all so random. Now they just ask the AI, which is trained on millions of genomes from all domains of life. It generated and evaluated thousands of different genomes, and narrowed the list to 300. The team synthesized these 300 different viruses in the lab and tested them to find 16 entirely new viable viruses. That’s never been done before.
AI doesn’t merely accelerate evolution. It proposed a genome on the other side of an evolutionary gap without nature having to walk through all the intervening generations.
In theory, if they dumped these new viruses into people, swamps or chicken pens which had the right E.Coli, they might peter out in a day, infecting a few bacteria and vanishing. But it’s also possible it could eventually spread far and wide — a right nightmare for E.Coli. These new viruses could live on and spread in perpetuity, killing a lot of EColi along the way until the bacterium evolved some resistance, after which the virus would mutate to counter that.
I asked one AI if this was correct and it replied at length expanding on it, until seconds later it suddenly deleted its own answer and said “This content can’t be shown”. We’re especially careful with request6s involving biological research and applications that could pose safety risks.” Well that sure makes me feel safe…
The big advantage here was not just the speed with which AI could create mutations, but that these mutations were fundamentally a leap forward in viral evolution that no amount of lab propagation was going to randomly generate. The phage called ΦX174 is only about 5,400 bases long, yet it contains 11 overlapping/interacting genes. The language model being used, called Evo, was trained on DNA as a language. It was trained on nine trillion nucleotides, and it found patterns and combinations that were unknown before. Evo was also specifically not trained on viruses that infect plants, animals, or people, and the safeguard appeared to work. Evo did not perform well on human-virus genomes. But the principle remains — if someone wanted to train an AI on nine trillion bases of mammalian viruses, the data is out there. We now know an AI can design a totally new viral genome.
In the next world war, a bad global citizen country could aquire the capability to design viruses and maybe also the drugs or vaccines to save themselves, before unleashing them onto adversaries.
Though as well all know, it’s much harder to design and test a useful vaccine.
The truth is we’ve always had deadly diseases to throw at any enemies — like Smallpox, Ebola, Marberg and Nipah. The thing that stopped us (apart from the morality bomb) is the blowback. Whatever virus is unleashed on the world will come back to hit the instigator too. This is where AI makes it all so much more dangerous. What if some overconfident General thinks his group has a vaccine that will save them? Suddenly the unthinkable becomes a strategy.
I would like the pull the pin on AI and just say “stop” — but that isn’t an option. The big news today was that US Intelligence agencies were warning that six Chinese Tech companies have been launching thousands of aggressive and targeted attacks, using more than 20,000 fraudulent accounts, to extract intellectual property from US leading AI systems.
“We can’t pause,” Bessent said. “You can’t, because the Chinese won’t pause.”
On the same day, the National Security Agency, Cybersecurity and Infrastructure Security Agency, and FBI issued a joint advisory accusing six China-based AI companies of conducting industrial-scale campaigns to extract capabilities from U.S. frontier AI models.
The agencies named DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun, and Z.AI, saying the companies had extracted billions of tokens through millions of requests to frontier models including Claude, GPT, Gemini, and Grok since at least late 2024.
Knowledge distillation is a machine learning technique in which one model is trained using the outputs of another. The technique is commonly used legitimately, including to develop smaller or less expensive models.
The advisory alleged, however, that the six companies engaged in “aggressive, malicious, and targeted” distillation at an industrial scale to extract restricted proprietary functions and capabilities from U.S. models. The NSA said the companies were systematically extracting U.S. model capabilities to train their own systems.
Beijing rejected the allegations.
The AI arms race is a life and death thing.
This article originally appeared at JoNova