3 stories, chosen and edited by hand.
01
OpenAI posted on X Friday defending its firing of safety researchers Jasmine Wang, Tomek Korbak and Mikita Balesni, saying an internal investigation found they violated policies on handling sensitive information — a "significant breach of trust." The post responded to an open letter the trio published Thursday, arguing they were actually dismissed for raising safety concerns and had acted within OpenAI's mission and norms.
My readOpenAI says there are breaches "beyond what's outlined in the letter" but won't say what they are, which is exactly the kind of non-answer that makes people assume the worst. When you fire safety researchers and refuse to show your work, you don't get to also claim the moral high ground on transparency. I'll be watching whether any actual details ever surface, because right now this is a standoff between two unverifiable stories.
Read the original at The Verge AI →
02
Cloudflare released Clef, open-weight decision models in 9B (Clef-Flash) and 27B sizes, that take a state and schema of typed questions and return per-option probabilities instead of generated text. Clef-Flash hits 38.8ms median latency versus 209.3ms for the 27B model, and the API is compatible with Typesafe AI's Jev System One. Weights are on Hugging Face, with hosting via Workers AI and a fine-tuning service coming.
My readI like that this skips the parse-the-JSON-out-of-an-LLM dance for routing decisions, and 38.8ms for Clef-Flash is genuinely useful if you're putting this in an agent's hot path. But the Hacker News skepticism is fair: this is classification with a vision encoder and a bigger context window than Jev, not a new category. The real test, like bugra_sa says on Reddit, isn't the benchmark number, it's whether Clef knows when to punt instead of staying falsely confident. I'll believe the latency claims when someone outside Cloudflare reproduces them.”} <br>``` Wait this needs to be pure JSON only. <br>Let me redo without stray text. <br> <br> <br> <br> <br> <br> <br> <br> <br> <br> <br> <br> <br> <br> <br> <br> <br> <br> <br> <br> <br> <br> <br> <br> <br> <br> <br> <br> <br>```json {
Read the original at InfoQ AI/ML →
03
Ben Affleck went viral this week for fluently discussing neural networks, transformers, tensors, and GPUs across interviews with GQ's Zach Baron and Bloomberg's Lucas Shaw at the Screentime 2026 conference. He sold his AI filmmaking startup to Netflix for a reported $587 million (he disputes the figure) and used fine-tuned open models in post-production on his movie "Animals."
My readWhat struck me isn't that a celebrity knows some jargon, it's that Affleck actually gets the distinction between training code and inference code, and why unfreezing weights on an open model to fine-tune toward "discrete tasks" is different from just prompting a chatbot. That's practitioner-level understanding, not talking points. The real story under the clip is a working method: build your own dataset to avoid likeness disputes, fine-tune, keep the proprietary layer. That's a template other studios will copy.
Read the original at TechCrunch AI →