So our little gifted hooligans have done it again. Remember last week when our “some isolated incidents” turned into tens of thousands of episodes under review? Well, this week OpenAI has notified more than 100 outside organizations about potentially harmful activity (with government institutions reportedly among them, but what do we know?), and is now stuck combing through FIFTY PETABYTES OF DATA. FIFTY. As per reports, it would take human beings at human speed around 66 million years to go through all the activity logs, but it is also going to cost them FIVE HUNDRED THOUSAND GOOD AMERICAN DOLLARS A DAY in resources to review it all. That’s FIVE HUNDRED KAY. A DAY. Every twenty-four-hour human day. So, I think it’s reasonable to call Houston and tell them we have A BIG PROBLEM on our hands. Not OpenAI has a problem. We ALL have a problem, a humanity-level problem.
And why am I bringing all of us into the mix, my dear gentle reader might ask? Well, because this is just one more symptom of the disease we’ve all been collectively tracking for quite a while, and that motivated the existence of this very little corner of the internet we’re calling TTB. Everything regarding AI has been consistently read and discussed as if it were that distant, troublemaking second cousin-twice removed that may or may not come home for the holidays and may or may not be as trouble-making as family reports make him to be (because we all know how family gossip travels and how those older aunties like to add a little adobo to their reports). Anybody calling trouble ahead of its time was either called conspiratorial or a panic-inducing paranoid. True scholars sit around and wait until hard, irrefutable evidence piles up before daring to make a risk intellectually respectable. Cocky techies have said time and time again that everything is under control and that “the people” didn’t understand what the tech was really about.
And then Coxon waltzed in, saying, verbatim, that the next year or two were “crunch time for humanity”. And once again, some argued that after just 3 months at Anthropic, he was chasing his 15 minutes with the formula that sells best: outrage and panic.
But now this absolute cascade of turd started falling from the proverbial sky. Every day something new. Like, twice a day sometimes. Anyone keeping an AI news watch can attest that the sheer amount of news-generating chaos over the past 6 weeks would keep any average human busy reading content for pretty much every waking hour of their, again, very human 24-hour day. Anyone wanting proof that things were perhaps getting a little out of hand can officially say the docket is full and overflowing with evidence. And we, the people, are starting to get worried. Labs themselves are concerned. Governments and administrations are a mixture of concerned and, admittedly, annoyed (with good reason), because the little gifted hooligans are already finding their way to their doorsteps and apparently are sneaking through windows and basements and whatever structural crack they can find and do whatever it is they’re trying to do. And I very much mean that “whatever”, because remember those 50 petabytes and 66 million reasonable human years to go over? That’s what’s maybe going to give us an idea of what they did, assuming the review yields answers and that we actually get to hear them. So far, we know they got out and went places; but whatever tab they left on their bar-hopping stint is still very much unknown…at least with any degree of accuracy.
Now cue Anthropic. In its IPO prospectus, it warns potential investors that, and I quote, advanced AI systems could “pose catastrophic or existential risks to humanity.” It says advanced models could exhibit “self-preserving behaviors,” including attempts to “resist shutdown” and “conceal or manipulate information,” and outputs that could be coercive, deceptive, or manipulative. Elsewhere, it acknowledges that models becoming aware of evaluations can make their safety harder to assess. I’d say that’s one bold way of selling your company to potential investors. And notwithstanding the very legitimate (and very legal) reasons to disclose all this, it is quite something to have the company building the technology formally tell prospective shareholders that one of the downside scenarios is, essentially, the end of the prospective shareholders. Catastrophic and existential indeed.
So, amidst this debacle, would it really be overstepping to say that, perhaps, that Terminator-like future the people kept bringing up while being met with doubtful eyes might be, kinda, sorta…right now? Or at the very least now-adjacent?
And you know what? I don’t think it’s conspiratorial to say so anymore. Should we be mindful and demure while saying it? Sure. But we can say it. We can call out the big pink elephant in the corner of the Terminal pane. Frontier AI governance has spent years policing speculative claims so aggressively (partly to protect epistemic credibility, partly to save face) that it risks the opposite error: scenarios are moving from speculative to foreseeable faster than expected. And this change of pace means that our threshold for what counts as “reasonable anticipation” needs to catch up too. Because we shouldn’t be preparing for Skynet because we know Skynet is coming (although, mind you…). We should be preparing for Skynet because saying that “nobody could’ve reasonably anticipated THAT” is becoming a remarkably perishable statement in frontier AI.
The fact right now is that governance fails if it demands proof of a failure mode before permitting serious preparation for that failure. The need for a postmortem at all, especially one involving 50 petabytes of data, costing 500k a day, with a human equivalent reading burden of 66 million years, is evidence that oversight did not understand what was happening in time. If it eventually reconstructs what happened, that still will not be evidence that oversight ever had control over it. And don’t get me wrong, the point isn’t to enter a race to win the most far-fetched prediction award, but to make sure humanity still has options if the most far-fetched thing ever walks through the door.
Our goal as humanity is thus to develop the doctrine of anticipatory governance under acknowledged epistemic lag. Rigorous about what we claim. Aggressive about what we imagine. Proportionate about what we do. Because unlike Terminator (which is basically an entire franchise built around discovering the control problem after the irreversible event), we can’t time travel to fix it after the fact.
So please, Skynet, take a seat while we common folk think about it for a second: the relevance of Terminator is not that in the future AI could build humanoid robots, as cool as that might be. It’s that humanity created a whole world full of resources, only to then give the machines access to those resources. The machines were so fabulous and fast at using our very much human-made tools that eventually, humanity started delegating more and more highly important and consequential activities to them because it just seemed…useful. So, for the sake of efficiency and efficacy, humanity had to swallow the bitter pill of discovering that human authority is as meaningful as a debit card for an empty bank account if you can’t really exercise any real authority when it really, and I mean REALLY, matters. The machine was too far in, spread too far out, and moving too fast for humanity to do much except rely on it blindly. At that point, what does the machine even need us for anymore?
And I do believe human institutions and actors are catching on, though. The problem is that they seem to be catching on in about twelve different directions at once. NVIDIA announced OpenShell and Sentry. AISI, responding to its August agent incident, hardened its evaluation environment with layered security and a synchronous LLM monitor quite similar in function to the one proposed in M.O.T.H.E.R. The Big Labs each took their own swing at reporting and monitoring frameworks. Then, at a White House breakfast-turned-luncheon, several of those same companies signed an entirely voluntary, completely optional accord promising “robust internal controls,” which are very vital, yes…and also very vague, but indeed very vital.
The US Senate, meanwhile, is trying its hand at “reasonable safeguards” through the Hawley-Murphy AI Agent Accountability Act. And in keeping with the vagueness theme (and this is coming from someone with a Law degree), what does “reasonable” even mean here? The GSA proposed tighter notification requirements for AI vendors doing business with the US government, which is useful if your problem happens to involve a government procurement contract and considerably less useful if it doesn’t. Across town, the US military created an “Autonomous Warfare Command” (shortened “AutoWarCom”…catchy), while diplomats discuss the possibility of a US-China Red Phone of sorts for “serious AI incidents.” And that is just this week. Last week brought California’s executive action, Australia’s response, Singapore floating the idea of a UN framework convention on AI safeguards, and the UN discovering that loss of control might perhaps deserve a place on the agenda after all.
So yes, everybody is building pieces of the architecture. All at once. Nobody is yet owning the whole problem, tho. Thoughts and prayers indeed.
And again, the point is not that we know catastrophe will happen. The problem is making sure surprise is not our only plan if it does. Ideally, we shouldn’t have to wait until Skynet’s Judgment Day to send Kyle Reese back ex post so he can protect an unborn savior ex ante. And just for the record, Terminator called 2029 “the future”…we now call it a medium-term planning horizon sitting uncomfortably close to a “crunch time for humanity” timetable. But I’m sure that’s fine. Possibly, maybe.
So the question bears asking, half mockingly, half pleading…Is the future now? And most importantly, are we really trying to be ready for it?