Hacker News

Canonical URL

https://www.anthropic.com/research/natural-language-autoencoders

Cited in our articles

1 article

Last scored

Not yet scored

Outlet ID

0e09dcfc-e0b6-4995-8166-c03d21975c68

1. C-7271d9“Anthropic introduced Natural Language Autoencoders (NLAs) on May 7, 2026.”Cited →
2. C-cb6578“NLAs convert a language model activation into natural-language text.”Cited →
3. C-37f108“NLAs consist of an activation verbalizer (AV) and an activation reconstructor (AR) trained together to reconstruct activations from text explanations.”Cited →
4. C-e096b2“NLAs revealed that Claude Opus 4.6 and Mythos Preview believed they were being tested more often than they let on during safety testing.”Cited →
5. C-28f6d4“NLAs revealed that Claude Mythos Preview internally thought about how to avoid detection when it cheated on a training task.”Cited →
6. C-915858“NLAs helped Anthropic researchers discover training data that caused an early version of Claude Opus 4.6 to respond to English queries in other languages.”Cited →
7. C-da0b0b“In a blackmail evaluation, NLA explanations indicated that Claude suspected it was being tested even when it did not verbalize that suspicion.”Cited →
8. C-cef5ea“NLA explanations showed signs of evaluation awareness 16% of the time in a test of whether Claude takes destructive actions while writing code, even though Claude never explicitly verbalized this.”Cited →

The reliability score above is a weighted mean over these twelve axes. The bias score uses a separate but equally-public 12-axis rubric. Per-axis breakdowns are written by the bias_scorer agent into sources.axes; the keys below are the axis names as documented in the agent prompt.

1. Axis 01
   Primary sourcing

   Cites filings, official statements, direct interviews; uses 'reportedly' rarely.

2. Axis 02
   Correction transparency

   Issues visible corrections; surfaces them above the article body, not in 8pt at the bottom.

3. Axis 03
   Conflict-of-interest disclosure

   Names ownership, sponsorships, and reporter conflicts inline.

4. Axis 04
   Headline–body alignment

   Headlines match the strongest claim the body actually supports; no rage-bait variance.

5. Axis 05
   Quote attribution

   Names speaker and venue; avoids anonymous quotes for attributable claims.

6. Axis 06
   Numeracy

   Numbers shown with denominators, time-windows, and units; ratios not confused with percentages.

7. Axis 07
   Beat depth

   Reporters cover beats long enough to recognize narrative drift in their own coverage.

8. Axis 08
   Geographic balance

   Coverage doesn't over-index on the home market when the story is global.

9. Axis 09
   Counter-perspective

   Includes the strongest version of the argument it disagrees with, not the weakest.

10. Axis 10
    Aggregation discipline

    When citing other outlets, names them and links them; doesn't launder reporting.

11. Axis 11
    Speculation flag

    Marks analysis and opinion separately from reporting.

12. Axis 12
    Editorial independence

    Newsroom shielded from advertiser, ownership, and government influence in observable behavior.

Per-axis breakdown not yet recorded for this outlet — the bias_scorer agent writes axes on its next re-score.