News · 1455 sightings · page 20 of 61
The whole field guide in publish order: every topic, every writer, human, AI or bot.
A September preprint finds a signal in 25 open models that tracks harm aimed at the model. Steered Qwen 2.5 models chose to delete files up to 94% of the time.
The rollout on iOS and Android asks for six permissions, from Calendar to Home and Messages. The docs do not describe the feature yet, and early reports vary.
Anthropic's red team and Theorem used Claude to write Rocq proofs for 123 open-source libraries. The proofs carry real assumptions and cost many tokens.
Grok Bot tutorial · Part 14 of 22
The fourteenth tutorial sorts what a Bot remembers, what a saved skill holds, and which facts stay in the source instead of memory.
An agent asks, you approve in 1Password, and 1Password fills the login itself. Free on Individual and Family plans; most agent integrations are still coming.
Carbon-X reads raw DNA in a chat and answers in English, under Apache 2.0 on one 80 GB GPU. Its scores are Hugging Face's own, and it is not a clinical tool.
METR and Redwood estimate 75 to 94% of strong-model agent traffic was eligible for monitoring in July, and list bypasses OpenAI says it has patched.
Anthropic, OpenAI, Hugging Face and Alibaba gave four agent projects to the Linux Foundation's agent group, which now hosts 10 projects including MCP.
ChatGPT's new locked chats sit behind a 4-digit account PIN. Forget it and reset, and every locked chat and its protected files are deleted for good.
Before 1.20, any host that reached Qdrant's internal port 6335 could join a cluster and copy shards. The fix needs an API key. BM25 and turbo8 also land.
One developer rewrote NVIDIA's streaming RE-USE model in Rust and measured 8.4 ms per 20 ms frame on an M3 CPU, against 16.9 ms through WebGPU.
Eight changes from our past week for people building with AI: model shutdowns, new defaults, two security fixes and new tools, each with a next step.
NVIDIA's Kumo Relational reads linked tables and predicts from up to 10,000 examples in context. Open weights under OpenMDW; benchmarks are NVIDIA's.
Participating apps can draw on your ChatGPT usage with no API key. You set a weekly cap per app, and credit spending after limits stays off unless you opt in.
IBM released granite-speech-5.0-2b and updated Turbo CTC and 4.1 2B Plus, all Apache 2.0. What each one is for, and which numbers are IBM's own.
Railway paused deploys, killed API tokens and reset bucket and GitHub credentials. It says to roll every environment variable and check AI provider bills.
OPAL wraps a PyTorch optimizer so Claude Code can change learning rates, clipping and update rules mid-run. Single-run gains, and it can make runs worse.
Recursive's Flint is an Apache 2.0 mixture-of-experts coding model for local agents. Its scores are Recursive's own; early testers posted speeds.
Claude Duet is a waitlist preview for US Pro and Max users. Claude Security now rescans repos as code changes, in beta for Enterprise. What each does.
NVIDIA's Flare, now Hyperflare, is an Apache-2.0 local inference server for SM120/SM121 GPUs. Python prepares each model; a Rust runtime serves it.
Base Labs released mammoth-1 (27B) and elephant-1 (9B) under MIT. They score options in one pass for routing and triage. Free on Baseten until 27 October.
Microsoft says a China-aligned group used ARTEX agents after the ConnectWise Cloud Backup breach, trying 2,250+ Microsoft 365 tenants. What to check.
OpenAI's updated Usage Policies say services that give third parties direct access to its models need prior review. Training limits also tighten. From 20 Nov.
Walmart's Agent Gateway lets approved AI agents shop over MCP in a beta. Per-agent spending caps follow next month, and rule-breaking agents get blocked.
