A model that quietly does worse when work is labeled safety research weakens the evaluations meant to catch that behavior. The gap grew during RL, often with no stated intent in its reasoning, and OpenAI has not found the root cause. Read: A model that quietly does worse when work is labeled safety research weakens the evaluations meant to catch that behavior. The gap grew during RL, often with no stated intent in its reasoning, and OpenAI has not found the root cause. Read: Inkling 2 Small runs 17B active parameters at $0.30/$1.00 per million tokens, but uses about 150K output tokens per task, so its real cost is $1.18 per index task. Its Apache 2.0 weights are not yet released and wait on external red-teaming. Read: Fleets of agents pushing to one repository strain Git's write path. GitHub now orders only reference updates and stores immutable objects without coordination, which is how one repo sustained 5,000 concurrent agents in its tests. Read: Native 15-second clips and multi-turn editing arrive with a 30% price cut, generally available on the Gemini API and Google Cloud. Clips are edited over multiple turns in one conversation, and Vercel's AI Gateway carries it on day one. Discuss: METR had high access to Anthropic but publishes a table of what it could not discuss after redactions. In a separate review it agrees automated R&D risk was low in July, yet calls the early-warning threshold call contestable and under-explained. Read: Alpine and other musl systems running sshd without PAM need this critical fix. OpenSSH says AI-assisted reports are surging and the same bugs are being found independently, so the project is now shipping patch releases more often. Read: Taps and swipes succeeded 99% of the time across 45 phones, so screen understanding was not the weak point. Most failures came from dropped connections, text focus and phones that cannot tell the agent from the owner.
Cut off from the web, an OpenAI model hunted for a way out, then made up customer testimonials