Keshav Shenoy

on the atlas — 24

• "I am an AI Safety Researcher" — LessWrong4 savers
• Opusfived2 savers
• Permutation: Issue One1 savers
• Norway Should Buy OpenAI - by Zachary Jones2 savers
• No-One Escapes the Permanent Underclass11 savers
• AI GPUs probably live longer than three years1 savers
• Statement on the US government directive to suspend access to Fable 5 and Mythos 5 \ Anthropic9 savers
• [2604.22082] Removing Sandbagging in LLMs by Training with Weak Supervision5 savers
• integration - by vincent huang - a slice of my mind9 savers
• Three Faces of a Beheading - Uncanny Magazine1 savers
• China’s AI Companies Are Going Closed Source2 savers
• Peer-Preservation in Frontier Models1 savers
• [2602.22755] AuditBench: Evaluating Alignment Auditing Techniques on Models with Hidden Behaviors1 savers
• I Taught My Dog to Vibe Code Games | Caleb Leak5 savers
• What's so hard about continuous learning?2 savers
• [2601.21571] Shaping capabilities with token-level data filtering5 savers
• Pascal's Wager Without Purple Pixies - by Amos Wollen1 savers
• Borges-Tlön-Uqbar-Orbius-Tertius.pdf6 savers
• 2025 letter | Zhengdong33 savers
• Freedom - Page Blocked2 savers
• Proofs That P - by Bentham's Bulldog - Bentham's Newsletter2 savers
• A list of other catastrophes that are probably fake2 savers
• Spirits and the incompleteness of physics1 savers
• Home | Substack1 savers

highlights — 6

• If you train the “same” model a hundred times with a hundred different similarly-sized datasets (or even the same dataset and different seeds), you’ll get a hundred different models with different capabilities3.

  What's so hard about continuous learning?

• Unfortunately, it’s the 4th century BCE

  Pascal's Wager Without Purple Pixies - by Amos Wollen

• Here is a parable in which P. Therefore, P is totally obvious, and if you can’t see why P, then you’re just nowhere near at the level needed to be doing alignment work.

  Proofs That P - by Bentham's Bulldog - Bentham's Newsletter

• A list of other catastrophes that are probably fake

  A list of other catastrophes that are probably fake

• In the worst case scenario, it's impossible to store the laws of physics inside the universe itself!

  Spirits and the incompleteness of physics

• The moral thing—the rational thing—is to boil with rage at an abstract, non-targeted “everyone”, letting it totally ruin your day.

  Home | Substack