The Future of Everything Is Lies, I Guess: Safety

A bleak thesis, delivered plainly

New machine‑learning systems are not just buggy. They’re dangerous. Researchers warn that LLMs shift the economics of harm — lowering the cost for targeted fraud, harassment, psychological abuse, and even kinetic effects when models are paired with automation. The blog post argues that the so‑called “lethal trifecta” is actually a unifecta: LLMs simply cannot safely be given the power to fuck things up. Harsh? Yes. Credible? Increasingly so.

Alignment isn’t a magic spell

The kernel of the argument: “alignment” — the project of making models benevolent — is not baked into the math. It’s a product of expensive data work, human reviewers, and secondary models that police outputs. Do that work and you get a friendly model; skip it and you get... something else. It has been reported that state actors have long tried to exfiltrate IP from frontier labs, and those threats would only accelerate if the expertise isn’t locked down. OpenAI allegedly believes some competitors trained off their outputs; if true, that’s the textbook shortcut around expensive alignment work.

Four crumbling moats

Why is the bar falling? Four weak points, all eroding at once: access to training and inference hardware is expanding; the underlying math is public and the software know‑how is bleeding out of big labs; web corpora are trivially scraped at scale; and large pools of human raters are effectively replicable or bypassable by model‑generated labels. Put them together and the industry has lowered the barrier for building unaligned systems. So who gets to decide what’s “friendly” when anyone with funding can spin up a model?

A near future that pinches

The practical result is ugly and immediate. Expect more subtle, scalable scams, harder moderation problems, and content that’s easier to weaponize against people’s minds. Semi‑autonomous weapons are no longer a thought experiment. Moderators will carry more of the moral load. Can policy, engineering, and civic institutions keep pace? Maybe — but don’t bet on complacency. This isn’t sci‑fi; it’s a policy fight that starts with who gets the GPUs and ends with who pays for the cleanup. Not exactly the future we ordered.

Sources: aphyr.com, Hacker News