The Future of Everything Is Lies, I Guess: Safety
Concerns over large language models are shifting from abstract existential risk to concrete safety failures, power imbalances, and regulatory overreach. Commenters argue that current “alignment” mainly protects the interests of model owners while doing little to prevent fraud, harassment, or abuse at scale, and point to examples like UK sites geoblocking users rather than absorbing new legal liabilities. Others counter that AI is just another dual‑use technology like the internet: attempts to tightly control it may entrench big players, while the real challenge is designing systems and laws that mitigate escalating harms without sacrificing useful capabilities.
Geo-blocking and the UK Online Safety Act
- The article is unavailable in the UK due to the Online Safety Act; the author geoblocked the UK after regulator guidance targeting “one‑person” services that don’t do age checks or child‑risk assessments.
- A presentation reportedly said geoblocking would count as compliant, despite other statements downplaying it as sufficient.
- Some see widespread geoblocking as a good way to create pressure inside the UK; others are pessimistic that it will matter.
Series structure and Hacker News dynamics
- Several commenters note that posts from this domain reach the front page quickly due to long-standing reputation.
- The multi-part series has very skewed readership: intro and a few sections got heavy attention; others (including ones some found strongest) sank.
- Some suggest a clearer overarching abstract or table of contents might have signaled the breadth and balance better.
Alignment, evolution, and “niceness”
- Debate over whether human prosocial behavior is meaningfully different from LLM “alignment,” or just both results of optimization (evolution vs gradient descent).
- Some argue evolution also produces antisocial agents; nothing guarantees “niceness” toward humans.
- Others think prosocial behavior can be game-theoretically favored but remains hard and expensive to engineer in models, and capitalism may not incentivize it.
Risk, misuse, and security impacts
- Many agree LLMs lower the cost of sophisticated fraud, social engineering, and offensive security, especially for non-experts.
- Some emphasize that defenses can improve too; others counter that attacker–defender effects are asymmetric and ordinary users will bear more defensive burden.
- Jailbreaking is described as still easy, though via changing techniques; guardrails are seen as fragile “patches.”
Regulation, access, and power asymmetry
- One camp wants heavily controlled or registered models to curb criminal uses; another fears concentration of power in a few labs or governments.
- Some welcome the falling cost of training “unaligned” or minimally aligned models as a way to escape a small cartel’s values.
- Recurrent theme: alignment often means “aligned with whoever pays for the model,” not with end users, reinforcing existing SaaS-style power imbalances.
Optimism vs pessimism about AI and technology
- Critics of the article see it as demonizing technology and repeating internet-era pessimism; they argue harmful uses are inevitable and benefits (e.g., personal assistants, coding agents) are large.
- Others question whether the internet itself has been a net positive, citing surveillance, addiction, fraud, and disinformation as warnings for AI’s trajectory.
- Several stress that thoughtful, often critical scrutiny is necessary to avoid repeating past mistakes, not mere “luddism.”