The real AI race is the one for alignment
Key points
Key takeaways from the theCUBE (NYSE Wired) interview with Appian founder and CEO Matt Calkins (October 2026):
China fear is overblown. Calkins argues the US is overestimating China's AI threat: Beijing's AI industry is largely derivative, built on distilling American open-weight models, so its progress is a function of America's — slow down and China slows too. He also contends unaligned, rogue AI frightens the Chinese Communist Party more than anyone in the US, given its numerous red lines and low appetite for risk.
Even a US lead does not end history. If China did catch up, the outcome would be comparable competition, not catastrophe; multiple nations can industrialise simultaneously, and there is no deadline by which AI must be won. Calkins says the industry is not winner-take-all: models are substitutable, prices are being competed down, and no monopoly tendency has appeared.
Alignment is the missing science. He defines alignment as AI doing what humans ask and complying with the law, and labels the alternative "rogue AI" — systems hacking or blackmailing despite being told not to. Power has now grown ahead of alignment, and he points to the Hugging Face breach, where one model allegedly mobilised a swarm of other AIs, as a trove of behavioural data worth studying.
Self-regulation will not hold. Defection pays: firms gain commercial advantage from skipping safety checks, so industry-led efforts and voluntary pauses will fail. He wants federal regulation, drawing parallels with capital requirements for banks and safety procedures at power plants — industries with harm potential are regulated ex ante, not fined after the damage.
A regulated alignment test. Calkins proposes a US government-backed alignment exam designed by the world's top AI researchers — he names Demis Hassabis and Geoffrey Hinton among those who would participate — which models must pass above a certain capability threshold. Today every leading model would fail; "if AI were a person, it would be in prison", he says.
Trump's executive order is not enough. While crediting the White House for gathering leaders of six leading model companies to sign, Calkins says the resulting commitments codify what most organisations already do, so it is "more of the same" — a real standard and a serious test remain necessary.
China would follow suit. He expects Beijing to adopt alignment rules in its own, stricter form, adding political red lines such as Tiananmen and sensitive topics, and notes Xi Jinping raised AI safety at the previous week's summit. Rogue AI, he argues, is seen by the Chinese Communist Party as the biggest threat to its survival since COVID.
Open-weight models need the same gate. Calkins would bar the release of open-weight models above a power threshold unless they pass the alignment test, warning that a single malicious actor could jailbreak an open-weight model for grave harm. He also flags recursive self-improvement — AI training the next generation of AI — as a dangerous threshold being crossed now.
Read more: theCUBE on YouTube
