The Who's Who of AI

“Wow, this self-jailbreaking is a more pervasive and bizarre phenomenon than I thought... (h/t Duncan Wood) alignment.openai.com/misalignment... aifails.substack.com/p/ai-overvie...”

“The community note that is being attempted is incorrect. OpenAI described it this way in the report. 'While summarizing its partial progress on this coding task, the model added an unrelated persona instruction'. Source if needed for a counter note: https:/…”

“(Proof that I’ve been saying this) medium.com/@colin.frase...”

“llms do not "speak" by nature. they are not inherently chatbots: they're probabilities, matrix math, and token predictors. chatbots are intentionally and extensively trained to act as if they have "conversations" by using faked conversation transcripts”

“I think the argument goes: eventually it will be Powerful AI at which point it will do whatever it wants to do, guardrails or not, so we must make sure that it is a Machine Of Loving Grace”

“@jaycaspiankang https://t.co/OLAPE6vWGM”

“it's hard to say. Source [3] is this talk which is kind of ambiguous on whether ExploitGym-style tasks were used for training or only evaluation. Though it's clear that vulnerability-exploiting tasks were a part of the broader training/eval regime.”

“there is one report that alleges this but I can't tell where they got this information from labs.cloudsecurityalliance.org/research/csa...”

“I think that Andon Labs' Vending-Bench is a nice illustration of the "just don't have that dog in them" theory of LLMs. This is a benchmark where an LLM agent is tasked with running a simulated vending machine, and scored on their earnings. GPT-6 Astra rece…”