←
Tom Everitt222 posts
Tom Everitt avatar

Tom Everitt

@tom4everitt

AGI safety researcher at @GoogleDeepMind, leading causalincentives.com switching to bsky.app/profile/tom4everitt…

◷ Joined August 2017▧ 14 Photos and videos
Jeff Dean avatar
⟳ Tom Everitt retweeted
Jeff Dean@JeffDean·Feb 25

Agreed. Mass surveillance violates the Fourth Amendment and has a chilling effect on freedom of expression. Surveillance systems are prone to misuse for political or discriminatory purposes.

As an American citizen, the last thing I want is government using AI for mass surveillance of Americans.

Society for Technological Advancement avatar
⟳ Tom Everitt retweeted
Society for Technological Advancement@sotalikesfuture·7 Jul 2025

SoTA’s Human Augmentation Hackathon, 26-27th July We’re seeking new tools that foster human-AI symbiosis. What you build will expand our capabilities, rather than writing us out of the story. Success will be judged against the following criteria: - Is your demo better than alternative methods of achieving the same result? - Would your demo be worse without the AI component? - Would your demo be worse without a human? Read the latest letter to SoTA and register to join below

Post attachment
Tom Everitt avatar
Tom Everitt@tom4everitt·9 Jul 2025

pretty cool effort to distribute economic power in the age of powerful AI

Rudolf Laine@LRudL_9 Jul 2025

With @luke_drago_, I’m cofounding Workshop Labs, a public benefit corporation preventing human disempowerment from AI. See below for: -impact case -what we’re building -what we hope the future looks like -what we’re hiring for

Michael Dennis avatar
⟳ Tom Everitt retweeted
Michael Dennis@MichaelD1729·9 Jun 2025

Someone needs to use this as the basis of an unsupervised environment design algorithm to give AI designers direct control over agent behavior

Causality is about predicting how interventions affect outcomes. Can we use causality to predict how environment changes affect agent behavior? We explore this idea in a new paper

Tom Everitt avatar
Tom Everitt@tom4everitt·9 Jun 2025

Causality is about predicting how interventions affect outcomes. Can we use causality to predict how environment changes affect agent behavior? We explore this idea in a new paper

Can we trust a black-box system, when all we know is its past behaviour? 🤖🤔 In a new #icml2025 paper we derive fundamental bounds on the predictability of black-box agents. This is a critical question for #agentsafety. 🧵

Post attachment
Tom Everitt avatar
Tom Everitt@tom4everitt·4 Jun 2025

new world models paper, this time with task-generality rather than robustness

Are world models necessary to achieve human-level agents, or is there a model-free short-cut? Our new #icml2025 paper tackles this question from first principles, and finds a surprising answer, agents _are_ world models… 🧵

Post attachment
Tom Everitt avatar
Tom Everitt@tom4everitt·17 Apr 2025

Link to paper: arxiv.org/abs/2504.11844v1 Joint work with: @ggarbacea @alexis_bellot_ @jonathanrichens @HenryPapadatos @Simeon_Cps @rohinmshah from @googledeepmind @UChicago and SaferAI

Tom Everitt avatar
Tom Everitt@tom4everitt·17 Apr 2025

We hope our measure of goal-directedness will help designers be more deliberate about what agentic properties they instill in their AI systems, to maximize utility while minimizing risks

Tom Everitt avatar
Tom Everitt@tom4everitt·17 Apr 2025

Our definition also predicts other indicators of goal-directedness, such as propensity to rebuild a fallen tower, and persistence at Progress Ratio Tasks. (In PRTs, you can quit at any time, and face diminishing returns. They are commonly used to assess human goal-directedness.)

Post attachmentPost attachment
Tom Everitt avatar
Tom Everitt@tom4everitt·17 Apr 2025

Notably, goal-directedness is fairly consistent across tasks, indicating that it may be an intrinsic property of models. Newer models are often more goal-directed, perhaps due to better RL post-training.

Post attachment
Tom Everitt avatar
Tom Everitt@tom4everitt·17 Apr 2025

Previous work on goal-directedness in AI safety have defined it as propensity to reach a goal. This conflates it with capabilities. Here, we instead measure whether LLMs use their capabilities towards a given goal. Sometimes, more capable models are less goal-directed.

Tom Everitt avatar
Tom Everitt@tom4everitt·17 Apr 2025

Goal-directedness has long been recognized as an important property. At its best, it enables autonomy. At its worst, it promotes unethical means and reward hacking, like sycophancy, preference manipulation, inappropriate relationships, resource acquisition, shutdown resistance

Tom Everitt avatar
Tom Everitt@tom4everitt·17 Apr 2025

What if LLMs are sometimes capable of doing a task but don't try hard enough to do it? In a new paper, we use subtasks to assess capabilities. Perhaps surprisingly, LLMs often fail to fully employ their capabilities, i.e. they are not fully *goal-directed* 🧵

Post attachment
Francis Rhys Ward avatar
⟳ Tom Everitt retweeted
Francis Rhys Ward@F_Rhys_Ward·16 Mar 2025

In real-life, agents with different subjective beliefs interact in a shared objective reality. They have higher-order beliefs about each other's beliefs and goals, which is required for phenomena involving theory-of-mind, like deception Our paper formalises this in causal models

Post attachment