Uncensor any LLM with abliteration

Original Article ↗ Hacker News Discussion ↗

Interpretation of Asimov’s Three Laws

  • Long subthread debates whether the Three Laws were intended as parody, satire, or sincere optimism.
  • Consensus: “parody” is too strong; stories show the laws as flawed and tension-generating rather than purely positive.
  • Some argue you can’t assert authorial intent without direct evidence; others say artistic interpretation is inherently plural.

What “Abliteration” Does

  • Seen as a specific form of “representation engineering”: identify a “refusal direction” in internal activations and project it out.
  • Two modes discussed:
    • Inference-time intervention.
    • Permanent weight orthogonalization.
  • Parallels drawn to control/steering vectors; framed as “brain-chipping” models without full retraining.

Effectiveness and Model Quality

  • Mixed empirical reports:
    • Some users say abliterated Llama 3/Qwen2 behave like normal instruct models but without refusals, including on extreme content.
    • Others report “lobotomy”-like degradation: increased perplexity, broken stop tokens, self-talk, and lower overall quality.
  • One commenter with benchmarks claims negligible capability loss when done carefully; suggests some failures are implementation errors.

Open Weights, Safety, and PR

  • Strong view that this only matters when weights are downloadable; hosted APIs can’t be modified this way.
  • Several argue this is exactly why large vendors avoid releasing weights: once out, “safety” alignment can be stripped trivially.
  • Many emphasize that corporate “safety” is mostly brand and legal risk management, not global risk minimization.

Censorship, Free Speech, and Use Policy

  • Sharp split between:
    • Those who see uncensoring as obviously dangerous (bombs, bioweapons, CSAM-like content, election manipulation, harassment at scale).
    • Those who see LLMs as glorified search/BS generators and think censorship is paternalistic, inconsistent with free-speech norms, and largely ineffective given Google/books.
  • Meta’s Llama 3 license and its prohibition on “allowing misuse” is discussed; some treat it as a real contractual risk, others as practically unenforceable.
  • Some argue models trained on public web data have weak moral authority to impose strong downstream ToS.

Jailbreaks vs Weight Editing

  • Multiple people note prompt-based jailbreaks already bypass refusals (e.g., “legal department” or safety-framed prompts), so abliteration mainly lowers the barrier.
  • Others counter that making uncensoring require technical skill is still useful harm reduction; current safety is “too easy to jailbreak.”

Wider LLM Safety Frictions

  • Many concrete frustrations with overbroad refusals (regex for slur filtering, AWS/Gemini refusing basic auth or nuclear engineering questions).
  • Some argue safety should be implemented as a separate output filter, not baked into the core model.
  • Concern raised that as models become central infrastructure (education, policy, finance), having a few corporations define global speech norms is itself a major risk.