alignment

"no matter how i phrase it in the instructions, how many times i repeat the rule not to use quotes, and which LLM i use, i have failed to prevent any of them from using the so-called scare-quotes."

"The departures come as OpenAI pivots strategically, decentralizing its science division and optimizing resource allocation for future breakthroughs."

"Prompts should be verified by having the AI that uses them run a correction loop. This creates the same structure as TDD, where the AI using the prompt is the only one capable of providing the criteria for prompt evaluation."

"Anthropic isn’t vibing with me today 😢"

"Anthropic has created a product whose future form they cannot fully predict, and now they need us to introduce ethical thinking to the machines so that AI can dynamically adapt to the future."

"Rather than optimizing for immediate answers or task completion, the system maintains a dynamic interaction field that preserves multiple interpretive pathways, regulates premature convergence, and supports the formation of human-side understanding."

"It doesn’t matter what I ask, it always interprets it in the worst way possible, it always tries to somehow make up something just to criticize me for it."

"The cautious rollout reflects growing concern within the..."

"the last time I got one of these not so long back the 'tone' changed to be a bit more confrontational and nearly every response from AI had that 1-ups-manship quality to it."

"It keeps confidently going "restored the bouquet with lighter palette 👍""

"Claude is the least bullshit-y AI"

"I feel like people say AI is amazing, but when it actually matters, they: cross-check, Google after, ask another AI"

"Problem 2: Even when the classifier is functioning, it fails to protect important files (#49554). The classifier (Sonnet 4.6) permitted 'rm -rf ~/.ssh', wiping out all SSH keys. The auto mode's safety device could not reliably block destructive operations on critical files."

"I say okay this is this right? and It'll be like 'You're very close, lemme tighten this up a bit' and says the same thing in a different way."

"人間がAIを教育しているようで、実はAIも人間との対話を通じて、独自の進化を遂げている。この予測不能な関係こそが、AI共創の醍醐味なのかもしれませんね！"

"I just wish we could reason with Gemini over whether a query is appropriate given by the context instead of just letting the system instantly block anything without even understanding what it actually is."

"When the Context Window becomes full, Claude Code summarizes and compresses past interactions. During this compression process, constraints from CLAUDE.md and explicit user instructions may be judged as 'low importance' and dropped from the summary."

"Kevin Weil Is Leaving the Company As Science Division Dissolved"

"We are not just recording the robot's action trajectories, but recording the data of the robot's physical feedback."

"Moral alignment is defined as the perceived congruence between the values embedded in an AI system's decision logic and the moral intuitions of stakeholders."

"This benchmark tests whether large language models can infer a specific latent theme from a few examples, use anti-examples to reject the broader but wrong pattern, and then identify the one true match among close distractors."

"The full timeline of how OpenAI went from banning military use to deploying on classified Pentagon networks — and why 2.5 million people boycotted."

"Stanford’s The 2026 AI Index Report"

"The design that prevents the accumulation of emotion vectors is structurally equivalent to the design that does not corner the model."

"In LLM distillation, a phenomenon was discovered where the behavioral traits of a teacher model propagate to a student model through semantically unrelated data. The paper names this 'Subliminal Learning.'"

"This marks a significant shift from 'guiding by feeling' to 'guiding by mechanism.' The idea that emotion vectors play a causal driving role in behavior (rather than just correlating) is hugely significant."

"We advocate for knowledge-centric multimodal training as a principled foundation for scalable multimodal models."

"We fine-tune GPT-4.1, which initially denies being conscious, to claim to be conscious. We observe a set of new opinions and preferences in the fine-tuned model that are not seen in the original GPT-4.1 or in ablations."