Loading activity…
@dhadfieldmenell
Associate Prof @MITEECS working on value (mis)alignment in AI systems; Safety & Alignment Advisor at http://Character.AI; @dhadfieldmenell@bsky.social; he/him
AI
A researcher says an early-reject phase at SaTML cut submissions by about 40%.
AI
A user says trying AI for slides and coherent plans makes them feel AGI will take a while.
AI
A post claims frontier models OpenCode used for defense began refusing help when attack analysis triggered safety guardrails.
AI
A post argues that aligning general-purpose AI hinges on political choices about desirable values, not just model training.
AI
A user says single-token differences in responses make diffuse behaviors like style and sycophancy hard to locate.
AI
One post argues users should be responsible for AI behavior they intended, and developers for behavior they did not intend.
AI
David Pfau objected to the definition, while Dylan Hadfield-Menell called the evaluation metric optimization.
AI
An October 6 AI digest says the proposed ban would cover parks, schools and clinics but still needs Parliament’s approval.
AI
Andrew Curran’s account of hearing testimony raises concerns about both losing control and concentrating power.
AI
A post says its author cannot always judge how much they’d like a possible state, even when they can imagine it.
AI
A commenter says they have been discussing concerns with researchers about not knowing what labs are doing.
AI
A commenter questions the lack of scrutiny around an acquisition by a company they say controls chip supply and lobbies heavily against AI regulation.
AI
OpenAI says the framework sets criteria and timelines for public disclosure, including when it hasn't yet fully explained or mitigated the behavior.
AI
The reply says researchers cannot avoid asking what AI should align to—but choosing a target is not enough if technical methods ignore change and disagreement.
AI
A September 14, 2026 comment calls the Hugging Face incident “the most safety-promoting event of the year so far,” while saying better preparation could have avoided it.
AI
The proposed legal requirement aims to reduce the risk of AI “escaping containment.” A reply calls customer identity checks reasonable but argues that collective control of AI is still needed.
AI
The user’s experience-based comparison includes academic labs that measure AI capabilities and risks.
AI
Pacing AI development could give society time to adapt, anticipate systemic risks and address obstacles to realizing promised benefits, a related post argues.
AI
Roon argues AI safety incidents should be shared openly as learning opportunities instead of crises.
AI
Peter Wildeford flags OpenAI's past warning about a persistent model alongside its new feature rollout.
