Deloitte is refunding part of a $440K govt report after using AI (GPT-40), which hallucinated fake court cases + citations. Lawmakers called it a “human intelligence problem.”
AI didn’t fail, oversight did.
What do you think? Image: theaifield | IG #AI
SAFETY
-
Deloitte AI Hallucinations Raise Oversight and Accountability Questions
By
–
-
Backdoors Create Cybersecurity Vulnerabilities Exploitable by Hackers
By
–
The technical consensus is clear: you can't create a backdoor that only lets the "good guys" in. However they're dressed up, these proposals create cybersecurity loopholes that hackers and hostile nations are eagerly waiting to exploit. 6/
-
Encryption scanning creates dangerous backdoors negating security purpose
By
–
The reality is that scanning any content before it is encrypted negates the very purpose of encryption and is a dangerous backdoor. No amount of tinkering or word games can change that inescapable reality. 5/
-
Public Moral Acceptance of AI Replacing Human Work by Occupation
By
–
Paper showing what human work the American public thinks is morally permissible to replace with AI. Surprisingly, people are already okay with AI doing 58% of occupations (if AI does it well/cheap). A floor of 12% of jobs (mostly caregiving) would be morally repugnant to replace
-
Intent and Sincerity as Core Criteria for Effective AI Agents
By
–
To certain extent. But intent to do their job sincerely is the most important criteria.
-
Google DeepMind Dreamer 4 Mines Minecraft Using World Models
By
–
Crazy paper from Google DeepMind
— alphaXiv (@askalphaxiv) 7 octobre 2025
Dreamer 4 can mine Minecraft diamonds without ever touching the real game
It trained entirely inside its own world model, kinda like an “imaginary world”, showing agents can tackle long horizon tasks safely by just learning from videos! pic.twitter.com/5cG9oYyt51Crazy paper from Google DeepMind Dreamer 4 can mine Minecraft diamonds without ever touching the real game It trained entirely inside its own world model, kinda like an “imaginary world”, showing agents can tackle long horizon tasks safely by just learning from videos!
-
Vibes Engineering: Self-Contradictory Approach to AI Design
By
–
I actually really like the mismatch between "vibes" and "engineering", it makes the combined term self-contradictory in a way that I find mischievous and (hopefully) sticky
-
Tesla Self-Driving V14 Rollout Accelerates with 14.2 and 14.3
By
–
Many more Tesla self-driving V14 improvements still to come, but wide rollout of V14.1 has begun. 14.2 rolls out in a few weeks and then 14.3 a few weeks later, depending on safety testing. There is so much change that we are carefully confirming each one.
-
AI Image Generation Artifacts: Quality Issues in Visual Output
By
–
Nope it's full of artefacts, check the tongue
-
OpenAI Dev Day reveals Apps, Agents, and major partnerships
By
–
Top stories in AI today: – OpenAI's Dev Day Apps, Agents, and more
– OpenAI, AMD ink compute partnership
– Build AI customer support with Agent Builder
– Anthropic’s Petri for AI safety auditing
– 4 new AI tools, community workflows, and more Read more: https://
therundown.ai/p/openais-bigg
est-dev-day-reveals
…