Research and Safety News

Soccer player

Chaplain

Federal employees

AI Safety Researchers

Kindergarten Teachers

Dancer

AI Resistance Leader

Judges

Nurse

Politicians

Gynecologist

The job of judging which propositions are true (analytic philosopher).

Firefighter

Novelist

Surgeon (autonomous)

Police officer

Project Manager

Teachers

Barbers

Lawyers

Dentists

Therapists

Chefs

Game developer

social media content creators

Fluffer

Construction workers

Computer programmers

Soldier

Personal trainers

Fast Food Cook

Doctors

Prompt engineers

Fast Food Cashier

Bus drivers

Robot designers / engineers

Photographer

4c) Generalisation splitting, posted EoY 2026

8a) Emergent misalignment and optimisers, >50 upvotes

8c) Emergent misalignment and optimisers, posted EoY 2026

3b) A taxonomy of selection hacking, >3 comments

5b) Weak-to-strong is a bad idea, >3 comments

6b) DCI Philosophy and Theory of Change, >3 comments

1c) AI Safety Research Futarchy Follow-up, posted EoY 2026

2b) Prediction markets for good, >3 comments

3a) A taxonomy of selection hacking, >50 upvotes

1b) AI Safety Research Futarchy Follow-up, >3 comments

2c) Prediction markets for good, posted EoY 2026

7b) DCI Concrete projects, >3 comments

5a) Weak-to-strong is a bad idea, >50 upvotes

2a) Prediction markets for good, >50 upvotes

7a) DCI Concrete projects, >50 upvotes

4b) Generalisation splitting, >3 comments

1a) AI Safety Research Futarchy Follow-up, >50 upvotes

6a) DCI Philosophy and Theory of Change, >50 upvotes

5c) Weak-to-strong is a bad idea, posted EoY 2026

3c) A taxonomy of selection hacking, posted EoY 2026

7c) DCI Concrete projects, posted EoY 2026

6c) DCI Philosophy and Theory of Change, posted EoY 2026

4a) Generalisation splitting, >50 upvotes

YES

8) At least one paper from the agenda gets published at a top conference or has a companion LessWrong post with >100 upvotes within 1 year

2) I reflectively endorse focussing on it 1 year later

7) At least one paper from the agenda gets published at a top conference or has a companion LessWrong post with >100 upvotes within 6 months

1) I reflectively endorse focussing on it 6 months later

3) I consider the work produced useful 6 months later

6) A prominent figure in AI safety publicly endorses the agenda or some research produced by it within 1 year

9) At least three papers from the agenda get published at top conferences or have companion LessWrong posts with >100 upvotes within 2 years

4) I consider the work produced useful 1 year later

YES

YES

YES

Goal Crystallisation

Post-training order and CoT Monitorability

Detection game

Exploring more metacognitive capabilities of LLMs

Salient features of self-models

Research sabotage dataset

Model Emulation

Model organisms resisting generalisation

YES

Other

Before humanity colonizes the universe, we must ensure that the future we would build is one worth living in.

Digital minds research is an important and neglected approach to AI safety.

Fun Fact: If you put “fun fact” before a completely made up statement, people are 69% more likely to believe it.

Past-you may have been a willing and enthusiastic sacrifice to present-you, and assuming you'll remain wiser, it was a worthwhile trade.

It's a good idea to buy lots of Microsoft stock right now

It's a good idea to short lots of Microsoft stock right now

Capitalism will collapse in 2026

If you sacrificed what you valued the most in order to survive, then from viewpoint of past-you you-present are already as good as dead

NO

YES