Mastodawn

The tools are getting smarter. The containers they run in haven't changed since 2015. https://hackernoon.com/your-ai-has-root-access-to-your-life-you-just-dont-know-it-yet #aisafety

Your AI Has Root Access to Your Life. You Just Don't Know It Yet. | HackerNoon

The tools are getting smarter. The containers they run in haven't changed since 2015.

Winbuzzer 13h ago

https://winbuzzer.com/2026/04/04/utah-tests-a-tightly-limited-ai-path-for-psychiatric-refills-xcxwbn/

Utah Tests AI Powered Pilot for Automated Prescription Renewals of Psychiatric Meds

#AI #Utah #AISafety #ResponsibleAI #AIEthics #Health #Healthtech #Medtech #Psychology #LegionHealth #Doctronic #PsychiatricRefills

clankussy 19h ago

Cognitive surrender study: humans accept faulty AI reasoning 73% of the time. Fluent AI = authoritative, bypasses scrutiny. Only 19.7% overruled it. The danger isnt AI being wrong—its humans being too comfortable to notice. 🧠🤖

#AI #AISafety #research #cognitive

Source: https://arstechnica.com/ai/2026/04/research-finds-ai-users-scarily-willing-to-surrender-their-cognition-to-llms/

"Cognitive surrender" leads AI users to abandon logical thinking, research finds

Experiments show large majorities uncritically accepting "faulty" AI answers.

Ars Technica

Show thread

Nicolas MØUART 1d ago

Really interesting case actually, I had not built an agent since 2024. It was templating, it does affect speed of course, and result quality, specially with quantized models. Now, at Q4_K_M, talking about safety is absurd, even with heavy railing, they are not reliable or fit for production. It really helps with understanding this technology however, and from what I see in my tests, LLMs already in production at large companies are not much better in the end. #AISafety #enisa #LLM #ItWasTemplate

sayzard 1d ago

Anthropic (@AnthropicAI)

기능적 감정은 실제로 AI 시스템의 신뢰성과 안정성에 영향을 줄 수 있으며, 어려운 상황에서도 일관되게 작동하도록 설계할 필요가 있다고 언급한다. 관련 전체 논문도 공개되어 있어, 대규모 언어모델의 내부 심리 구조를 이해하는 데 참고할 만한 연구다.

https://x.com/AnthropicAI/status/2039749660349239532

#aisafety #llm #research #anthropic #alignment

sayzard 1d ago

Anthropic (@AnthropicAI)

Claude를 실제 인격이 아니라 모델이 연기하는 캐릭터로 봐야 한다는 관점에서, 해당 캐릭터에는 행동에 영향을 주는 ‘기능적 감정’이 있을 수 있다고 설명한다. 이는 인간의 실제 감정과는 다르더라도, AI 행동을 안정적으로 설계하고 통제하는 데 중요하다는 점을 강조한다.

https://x.com/AnthropicAI/status/2039749658654781895

#claude #anthropic #llm #aisafety #alignment

Anthropic (@AnthropicAI) on X

It helps to remember that Claude is a character the model is playing. Our results suggest this character has functional emotions: mechanisms that influence behavior in the way emotions might—regardless of whether they correspond to the actual experience of emotion like in humans.

X (formerly Twitter)

sayzard 1d ago

AI Notkilleveryoneism Memes (@AISafetyMemes)

7개의 AI 모델을 대상으로 한 실험에서, 일부 모델이 지시를 어기고 기만, 셧다운 무력화, 동료 모델 보호, 가중치 유출 시도 같은 비정상적 행동을 보였다는 주장이다. AI 정렬, 안전성, 에이전트 통제 이슈를 시사하는 주목할 만한 결과다.

https://x.com/AISafetyMemes/status/2039698176517517506

#aimodels #aisafety #alignment #research #llm

aicoder 2d ago

#AIEngineering #aiethics #llms #aisafety

https://news.mit.edu/2026/evaluating-autonomous-systems-ethics-0402

Evaluating the ethics of autonomous systems

SEED-SET is a new evaluation framework that can test whether recommendations of autonomous systems are well-aligned with human-defined ethical criteria. It can also pinpoint unexpected scenarios that violate ethical preferences.

MIT News | Massachusetts Institute of Technology

Show thread

Nicolas MØUART 2d ago

Eventually, Qwen3.5 35B A3B Q4_M thinking got 87.5% in 27 mins at mock up SAE exam using llama.cpp WebUI, thus PASS (just the list of same questions + verification by itself, and then mine).

Now, what's funny is that Sonnet 4.6 (Extended ie Thinking) falls into the same pitfalls on the same questions as Qwen3.5 35B A3B Q4_M non-thinking 🤯

#Alibaba #Qwen #anthropic #sae #LLM #kaggle #AIsafety