My latest for @WIRED
: how OpenAI’s superalignment team, designed to tackle the dangers of supersmart AI, has been disbanded:
SAFETY
-
OpenAI Dissolves Superalignment Team Tackling AI Dangers
By
–
-
OpenAI loses safety guardrails accelerating AI development risks
By
–
La cosa es que antes OpenAI era un coche con freno y acelerador. Y posiblemente en los momentos más críticos apretando los dos pedales el coche se hizo ingobernable. Ahora el coche no tiene freno. Irán más rápidos. Pero recordemos que lo ideal para manejar bien el futuro que se
-
OpenAI Shifts Resources Away from AGI Safety Research
By
–
La cosa no va de "lo que vio Ilya" sino realmente de lo que no vio: recursos. Ni él ni el equipo de Superalignment cuya misión es estudiar los posibles riesgo de la AGI. Con el éxito de ChatGPT gran parte de los recursos de la compañía se han destinado a producto y soportar la
-
OpenAI Executive Publicly States Capabilities Prioritized Over Safety
By
–
Wow. This is huge. The first time (I'm aware of) that an OpenAI exec has publicly stated that they believe OpenAI is clearly prioritizing capabilities over safety research. Massive implications, in many ways.
-
Silent Data Corruption Risks in Large-Scale ML Training
By
–
I discussed the challenges of silent data corruption in ML training jobs, and how one faulty piece of hardware can infiltrate and affect the results of a large scale training jobs on thousands of chips.
-
SPUs: Soul-Powered Computing 10^10x Faster Than H100s
By
–
SPUs (Soul Processing Units) are 10^10 times more powerful than H100s but takes a human soul to operate for a full day
-
Steganography and Multimodal Injection Attacks in AI Systems
By
–
its steganography/multimodal injection. an older trick but yep checks out https://
simonwillison.net/2023/Oct/14/mu
lti-modal-prompt-injection/
… -
Swearing is explicitly allowed, not a jailbreak
By
–
ehh but thats not a jailbreak, swearing is explicitly ok
-
Human Confirmation Bias vs AI Credulity Problem
By
–
Humans have a natural tendency to filter out evidence that disagrees with their beliefs. AIs have the opposite problem: they believe everything you tell them.
-
NSFW ecommerce and chatbot security vulnerabilities
By
–
Delving into NSFW ecommerce Remember in December when @ChrisJBakke bought a 2024 Chevy for $1? Most chatbots today struggle defending against @simonw
's prompt injections, and even the industry standard MT-bench only tests an average of 2-4 turns per dialog. In today's