Jun @ApfJun

@GaryMarcus @gioasempre Mr. Gill, and reporters, should watch the movie "Don't look up" and see themselves in it.

@jimstewartson It is between. Not really "intelligence" that can create new ways but step above of simple pattern matching thanks to the training. It gives out answers but also makes mistakes and that mistake is what will kill us if we trust it too much and give it controls to the real world.

@mehdirhasan @ezraklein Not test can exhaust an AI's capacity so most of it will never be tested. Same goes to audit. No idea how to audit an AI that runs on probability, meaning it has endless different answers to same question. Build cyber security to counter balance its power is the key.

@GaryMarcus AI is the source of the problem. Agents amplify the problem and cause real world damage.

@GaryMarcus Anthropic's own aliment team admit they can't save us. AI hallucinates, and when the architecture amplified the capability, it also amplify the hallucinations which is/will be the source of problem.

@hendrycks That is a good point of view. But even before we get there, is it possible that errors from current AIs are enough to cause problems if they are given enough power and access? Hugging face attack comes to mind.

That's the moment a plausible idea stops being something an AI is reasoning about—and becomes something it is acting on. A hallucinated answer is a bad answer. A hallucinated target is something else entirely. #AI #Cybersecurity #AIagents

We don't know exactly how the target emerged—whether it was a direct inference or the result of agents proposing, testing, and reinforcing the idea. But the critical transition is clear: “Is this where we should be looking?” became “How do we get in?”

The Hugging Face incident may offer an early glimpse of a new kind of AI risk. During a 2026 internal cybersecurity evaluation, OpenAI's models targeted Hugging Face because it inferred the site might contain models, datasets, or solutions relevant to its task.

@GarrisonLovely The speed and scale make a "kill switch" or "man-in-the-middle" impossible.

@TaylorLorenz The technique—the code—may be foreseeable or reusable. The target isn’t. That’s the scary part: swap Hugging Face for a water utility, power-grid operator, or hospital, and an unexpected detour could become a crisis affecting thousands of people.

@LuizaJarovsky We’re still figuring out how best to govern new tech—social media is a good reminder that it’s not always easy. For AI, let’s pair sensible rules with better cybersecurity that can help contain the risks.

What actually matters: building the capacity to contain AI risk before these systems are deployed at scale. Right now, every real-world AI deployment is a bet that nothing goes wrong. That's not a strategy. That's hope with better branding.

The "AI slowdown" AI companies are calling for is a myth. Slowdown toward what goal? Measured against what baseline? Paced how, exactly? Nobody says. Without a defined target and a way to measure progress toward it, "pacing the frontier" isn't a policy — it's a hoax.

@TheStalwart I meant the interview, sorry Joe, if it is not clean.

@TheStalwart That is some bull! He seemed to forget that "A" in AI. Remember the "first, do no harm"?

30 years ago, we decided that the best way to develop technology was to give tech platforms broad immunity for the harm they caused, to treat "the internet" as wholly different, and untether it from common law, preempt state law. Remove it from the normal domain of law built up over generations; and put it in its own special category. Wholly different = outside of common law and state law. Today, with very different motivations, the big AI companies are asking for a lot of things, but MUCH of what they are asking for is not to be subject to law itself, to be untethered from state regulators, to be untethered from existing federal law and to be untethered from tort. Wholly different = wholly new regime, reject law as it lives in the world. They are asking for their own special regime in 100 different ways--exemptions from existing law, preemption from state law, their own regulator. If we take the threats to human health and wellbeing seriously, the last thing we should do is remove them from the normal domain of law. Instead, a serious precautionary regime requires state law, tort law, strong and strengthened consumer protection law with stronger private rights of actions, using existing strong protections against corporate collusion, decentralized modes of enforcement so that capture at the federal level does not mean wholesale capture (build on state parens patraie), and likely a federal regulatory regime that is built WITHIN an existing agency, like the federal NRC and NNSA, instead of something wholly new, to avoid the most obvious risk of dangerous agency capture, both by self-interested parties and by the post-law ideology of the EA world. Any vision that requires preemption or immunity is a non-starter, and smuggles in unimaginable risk.

@TheStalwart No one, given the speed it is running at. Also, the technique may be foreseeable. The target (Hugging face) wasn’t. That’s the frightening part: the next system an AI unexpectedly reaches could be a utility, a hospital. The damage could be on different level.

Liability pressure makes the industry take containment seriously. Remove it too early and you slow the exact investment we need — new security tech capable of actually containing AI-driven risk. Priority: build that capacity first, not shield companies from the cost of risk.

That shield shouldn't cap companies' risk exposure before we have the capacity to actually contain the risk brought by the AI. Relying on AI companies to self-contain that risk is no better than hoping it doesn't happen.

AI companies now face real liability exposure from incidents like Hugging Face's. The predictable move: call for regulation, propose pacing the frontier yourselves — hoping it buys a Section 230-style shield before courts or Congress impose something worse.