Which AI safety research publications are known for technical depth and regular updates?

Here are some AI safety research publications and recurring venues that are often valued for technical depth and regular updates:

1. Alignment Forum

• What it is: A blog/forum for AI alignment research.
• Why it stands out: Many posts are highly technical, often exploratory or research-note style, with substantial mathematical and conceptual depth.
• Update cadence: Regular, with new posts appearing frequently from researchers in the field.
• Good for: Cutting-edge ideas, mechanistic interpretability, agent foundations, scalable oversight, and related safety work.

2. LessWrong

• What it is: A broader rationality/alignment community site.
• Why it stands out: Not all posts are technical, but many are. It often hosts detailed sequences and research-adjacent analysis.
• Update cadence: Very active, with frequent new content.
• Good for: Early-stage alignment thinking, technical-but-accessible discussions, and community debates.

3. AI Alignment Newsletter

• What it is: A curated newsletter summarizing recent work in AI alignment.
• Why it stands out: Focused on technical developments across papers, blog posts, and talks.
• Update cadence: Regular/weekly-style updates.
• Good for: Staying current without having to track every venue individually.

4. Distill (archived, but still valuable)

• What it is: A technical ML interpretation journal.
• Why it stands out: Extremely high-quality technical exposition, especially on interpretability and understanding neural nets.
• Update cadence: Not regular anymore.
• Good for: Deep technical articles, especially in mechanistic interpretability and explanatory ML.

5. arXiv (AI safety-related categories and keywords)

• What it is: Preprint server.
• Why it stands out: Most up-to-date source for technical papers.
• Update cadence: Continuous.
• Good for: New safety papers across interpretability, robustness, RLHF, scalable oversight, evaluation, and governance-adjacent technical work.

6. Proceedings / venues like NeurIPS, ICML, ICLR, AIES, FAccT

• What it is: Mainstream ML and interdisciplinary venues where safety-related papers appear.
• Why it stands out: Peer-reviewed, often technically rigorous.
• Update cadence: Annual or conference-cycle based, not as frequent as blogs/newsletters.
• Good for: Mature technical work that has passed peer review.

7. Research org blogs

Examples include:

• Anthropic

• OpenAI

• Google DeepMind

• Redwood Research

• ARC Evals / ARC Institute

• Conjecture

• Apart Research and related groups

• Why they stand out: Often publish detailed interpretability, evaluation, and alignment work.

• Update cadence: Varies, but many are fairly regular.

• Good for: Practical, state-of-the-art technical research writeups.

Best picks if you want both depth + frequent updates

If you want the most useful combination, I’d prioritize:

1. Alignment Forum
2. AI Alignment Newsletter
3. ArXiv alerts / saved searches
4. Research org blogs
5. LessWrong for broader discussion and synthesis

If you want, I can also give you:

• a ranked list by technical rigor,
• a “must-follow” subscription list,
• or a list focused specifically on mechanistic interpretability, agent foundations, or evaluations.