Top all-time · #1635
Alignment Science Blog
Company blog · alignment.anthropic.comaiscience
Alignment Science Blog is the company blog at alignment.anthropic.com, writing about ai and science. Since 2024, 45 different Hacker News readers have linked it 49 times, most often “Subliminal learning: Models transmit behaviors via hidden signals in data”. No feed was found, so read it on the site itself.
Too few sampled comments (26) to say how often HN recommends it.?
Why HN links it
HN readers link Anthropic's alignment research, especially the subliminal learning post showing traits can transfer through training on model outputs. Commenters cite it and related posts as support in AI discussions.
Summarised by Claude from a sample of 26 HN comments that link here. Each comment is marked recommending, critical or neutral (citing without judging); how this works.
Most-linked posts?
- Subliminal learning: Models transmit behaviors via hidden signals in data15 people · 17 mentions · first linked 2025-07-22
- The Hot Mess of AI3 people · 3 mentions · first linked 2026-02-03
- Teaching Claude Why3 people · 3 mentions · first linked 2026-05-08
- Automated Researchers Can Subtly Sandbag2 people · 2 mentions · first linked 2025-03-25
- Bloom: An open source tool for automated behavioral evaluations2 people · 2 mentions · first linked 2025-12-22
What Hacker News says?
I wonder if they know about this, basically training on LLM output can transmit information or characteristics not explicitly included https://alignment.anthropic.com/2025/subliminal-learning/ I’m curious, they have the example of raw base model output; when LLMs were first identified as zero shot chatbots there was usually a prompt like “A conversation between a person and a helpful assistant” that preceded the cha…
Comment linking /2025/subliminal-learning/
on “History LLMs: Models trained exclusively on pre-1913 texts”Submitted Subliminal learning: Models transmit behaviors via hidden signals in data as a story · 208 points
HN discussionThe difference is that the worker you hire would be a human being and not a large matrix multiplication that had parameters optimized by a a gradient descent process and embeds concepts in a higher dimensional vector space that results in all sorts of weird things like subliminal learning ( https://alignment.anthropic.com/2025/subliminal-learning/ ). It's not a human intelligence - it's a totally different thing, so…
Comment linking /2025/subliminal-learning/
on “Mistral releases Devstral2 and Mistral Vibe CLI”Submitted The Hot Mess of AI as a story · 14 points
HN discussion> Is it useful? Not really. So interestingly enough Anthropic has been using models of this kind for their alignment evaluations: https://alignment.anthropic.com/2026/reward-seeker/#:~:text=...
Comment linking /2026/reward-seeker/
on “WorldTUI: I trained a 360M model to pretend to be a terminal”Submitted Findings from a Pilot Anthropic–OpenAI Alignment Evaluation Exercise as a story · 4 points
HN discussionSubliminal learning: https://alignment.anthropic.com/2025/subliminal-learning/
Comment linking /2025/subliminal-learning/
on “Tell HN: X is opening any tweet link in a webview whether you press it or not”Submitted Reducing Reward Hacking by Localizing Bad Behavior Instead of Penalizing It as a story · 3 points
HN discussionJust a reminder that AI models' actions are reflections of the text humans write and the more we fret and make up doomsday scenarios that we then post online, the more likely a model is to do those things. https://alignment.anthropic.com/2026/teaching-claude-why/
Comment linking /2026/teaching-claude-why/
on “Ask HN: Why were OpenAI, Claude, and Grok simultaneously down?”Submitted Anthropic: Introducing The Conceptual Reasoning Index as a story · 3 points
HN discussionThey are talking to each other already. Whether for betterment remains to be seen. 2 Opus 4 talking directly to each other: https://www.iflscience.com/the-spiritual-bliss-attractor-som... "Learning" passing between derivatives of the same base models: https://alignment.anthropic.com/2025/subliminal-learning/
Comment linking /2025/subliminal-learning/
on “Is it a bubble?”Submitted Anthropic's Pilot Sabotage Risk Report as a story · 3 points
HN discussionThere was a recent paper that showed you can spread model’s behavior through training on outputs, even if you don’t directly include obvious markers of the behavior. It’s totally plausible that training off Claude’s outputs subtly affected GLM into mentioning “Claude” even if they don’t include the direct tokens very often. https://alignment.anthropic.com/2025/subliminal-learning/
Comment linking /2025/subliminal-learning/
on “GLM-4.5: Reasoning, Coding, and Agentic Abililties”Submitted Automated Subtle Emergent Misalignment Research as a story · 2 points
HN discussionSorry, I commented on the wrong article. I meant to post this under: https://alignment.anthropic.com/2025/subliminal-learning/ Regarding your comment, yes, it's well known in the ML world that machines are way better than humans at picking up on correlations. In other words, the output of a model can carry traces of its internal state, so if another model is trained on those outputs, it can end up learning the patte…
Comment linking /2025/subliminal-learning/
on “Distillation makes AI models smaller and cheaper”Submitted The Persona Selection Model: Why AI Assistants Might Behave Like Humans as a story · 2 points
HN discussionAh you are right, I was commenting on this article: https://alignment.anthropic.com/2025/subliminal-learning/
Comment linking /2025/subliminal-learning/
on “Distillation makes AI models smaller and cheaper”Submitted Bloom: An open source tool for automated behavioral evaluations as a story · 2 points
HN discussionI think Honesty can be evaluated. Does the model push back when it knows the user is wrong? How often does the model hallucinate data vs. say it doesn't know? Provide a prompt with contradictions or other issues and see if the model corrects you. Here is an article by Anthropic that explains what they do and mean in more detail: https://alignment.anthropic.com/2025/honesty-elicitation/
Comment linking /2025/honesty-elicitation/
on “Claude Opus 4.8”Submitted Bloom: An open source tool for automated behavioral evaluations as a story · 2 points
HN discussionA compiler is another thing whose honor and pride that the models have taken from the nerds. In the past, people would debate for hours about the “dragon book” v.s. “writing interpreters” and present their cool bespoke compilers in Show HN articles. Now models can produce 100,000 lines of code over two weeks with no human intervention that actually work and can compile significant project. Which way now nerd? The mo…
Comment linking /2026/hot-mess-of-ai/
on “We tasked Opus 4.6 using agent teams to build a C Compiler”Submitted Stress-testing model specs reveals character differences among language models as a story · 2 points
HN discussionhttps://arxiv.org/abs/2410.04332 https://alignment.anthropic.com/2025/selective-gradient-mask...
Comment linking /2025/selective-gradient-masking/
on “Meta Muse Glimmer – open weights 30B local coding model”Submitted Automated Researchers Can Subtly Sandbag as a story · 2 points
HN discussionWasn't it this one? Article: https://alignment.anthropic.com/2025/subliminal-learning/ Paper: https://arxiv.org/abs/2507.14805
Comment linking /2025/subliminal-learning/
on “LLMs consistently pick resumes they generate over ones by humans or other models”Submitted Automated Researchers Can Subtly Sandbag as a story · 2 points
HN discussionNo, “safeguards” goes beyond that, it’s a whole process at Anthropic: https://www.anthropic.com/news/building-safeguards-for-claud... https://alignment.anthropic.com/2025/introducing-safeguards-...
Comment linking /2025/introducing-safeguards-research-team/
on “Statement from Dario Amodei on Our Discussions with the Department of War”Submitted A Toy Evaluation of Inference Code Tampering as a story · 2 points
HN discussion> https://alignment.anthropic.com/2025/subliminal-learning/ Thanks, that's the research I was thinking about, but I couldn't recall the keyword to search for it.
Comment linking /2025/subliminal-learning/
on “The agent harness belongs outside the sandbox”Submitted Automated Researchers Can Reliably Mitigate Alignment Failures as a story · 1 points
HN discussion
